Deep Residual Learning for Image Recognition Kaiming He Xiangyu Zhang Shaoqing Ren Jian Sun Microsoft Research kahe, v-xiangz, v-shren, jiansun @microsoft.com - 中英文对照
translated: 2026-07-16
title: "Deep Residual Learning for Image Recognition Kaiming He Xiangyu Zhang Shaoqing Ren Jian Sun Microsoft Research kahe, v-xiangz, v-shren, jiansun @microsoft.com" aliases: - "ResNet" - "arXiv:1512.03385" source: "https://arxiv.org/abs/1512.03385" arxiv: "1512.03385" created: 2026-07-16 type: paper-translation status: translated tags: - paper - ml - deep-learning - vision
Deep Residual Learning for Image Recognition Kaiming He Xiangyu Zhang Shaoqing Ren Jian Sun Microsoft Research kahe, v-xiangz, v-shren, jiansun @microsoft.com - 中英文对照
中英文对照
<a id="S0001"></a> Source: p.1 S0001
Original: Deep Residual Learning for Image Recognition Kaiming He Xiangyu Zhang Shaoqing Ren Jian Sun Microsoft Research kahe, v-xiangz, v-shren, jiansun @microsoft.com { } Abstract 20 Deeper neural networks are more difficult to train.
中文: 为图像识别而深残留学习 Caiming He Xiangyu 张绍清 Ren Jian Sun Microsoft Research kahe, v-xiangz, v-shren, jiansun @microsoft.com {} 文摘 20 Deeper 神经网络更难训练.
<a id="S0002"></a> Source: p.1 S0002
Original: We present a residual learning framework to ease the training 10 of networks that are substantially deeper than those used previously.
中文: 我们提出了一个剩余学习框架,以方便培训10个比以前使用的网络要深得多的网络。
<a id="S0003"></a> Source: p.1 S0003
Original: We explicitly reformulate the layers as learning residual functions with reference to the layer inputs, in- 0 0 1 2 iter. 3 (1e4) 4 5 6 stead of learning unreferenced functions.
中文: 我们明确将各层重新划分为学习剩余功能,参照各层输入,在-0 1 2 里拉. 3 (1e4) 4 5 6 中取而代之的是学习无参考功能.
<a id="S0004"></a> Source: p.1 S0004
Original: We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth.
中文: 我们提供了全面的实证证据,表明这些剩余网络比较容易优化,并且能够从大幅度提高的深度中获得准确性。
<a id="S0005"></a> Source: p.1 S0005
Original: On the ImageNet dataset we evaluate residual nets with a depth of up to 152 layers—8 × deeper than VGG nets [41] but still having lower complexity.
中文: 在ImageNet的数据集上,我们评价了深度为152层的残余网——8-比VGG网更深[41],但复杂性仍然较低。
<a id="S0006"></a> Source: p.1 S0006
Original: An ensemble of these residual nets achieves 3.57% error on the ImageNet test set.
中文: 这些残留网的组合在ImageNet测试集上实现了3.57%的误差.
<a id="S0007"></a> Source: p.1 S0007
Original: This result won the 1st place on the ILSVRC 2015 classification task.
中文: 这一结果在ILSVRC2015分类任务上获得了第1名.
<a id="S0008"></a> Source: p.1 S0008
Original: We also present analysis on CIFAR-10 with 100 and 1000 layers.
中文: 我们还对CIFAR-10进行了100和1000层的分析。
<a id="S0009"></a> Source: p.1 S0009
Original: The depth of representations is of central importance for many visual recognition tasks.
中文: 表达的深度对于许多视觉识别任务至关重要.
<a id="S0010"></a> Source: p.1 S0010
Original: Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset.
中文: 仅仅由于我们极其深刻的表达, 我们得到了28%的相对改进 COCO对象检测数据集。
<a id="S0011"></a> Source: p.1 S0011
Original: Deep residual nets are foundations of our submissions to ILSVRC & COCO 2015 competitions1, where we also won the 1st places on the tasks of ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation. 1.
中文: 深残留网是我们提交ILSVRC & COCO 2015比赛1的基础,我们在该比赛中还获得了ImageNet检测,ImageNet本地化,COCO检测和COCO分化等任务的第1名. 1. 联合国
<a id="S0012"></a> Source: p.1 S0012
Original: Introduction Deep convolutional neural networks [22, 21] have led to a series of breakthroughs for image classification [21, 50, 40].
中文: 导 言 深革命神经网络[22,21]引出一系列图像分类突破[21,50,40].
<a id="S0013"></a> Source: p.1 S0013
Original: Deep networks naturally integrate low/mid/highlevel features [50] and classifiers in an end-to-end multilayer fashion, and the “levels” of features can be enriched by the number of stacked layers (depth).
中文: 深层网络自然将低/中/高等特征[50]和分级器以端到端多层的方式整合,而地物的"级别"可以通过堆叠层的数量来丰富(深度).
<a id="S0014"></a> Source: p.1 S0014
Original: Recent evidence [41, 44] reveals that network depth is of crucial importance, and the leading results [41, 44, 13, 16] on the challenging ImageNet dataset [36] all exploit “very deep” [41] models, with a depth of sixteen [41] to thirty [16].
中文: 最近的证据[41、44]显示,网络深度至关重要,具有挑战性的ImageNet数据集[36]的主要结果[41、44、13、16]都利用了“非常深”的[41]个模型,深度为16 [41]至30 [16]。
<a id="S0015"></a> Source: p.1 S0015
Original: Many other nontrivial visual recognition tasks [8, 12, 7, 32, 27] have also 1http://image-net.org/challenges/LSVRC/2015/ and http://mscoco.org/dataset/#detections-challenge2015. )%( rorre gniniart 20 10 00 1 2 3 4 5 6 iter. (1e4) )%( rorre tset 56-layer 20-layer 56-layer 20-layer Figure 1.
中文: 许多其他非三角视觉识别任务[8、12、7、32、27]还有1http://image-net.org/challenges/LSVRC/2015和http://mscoco.org/dataset/#detections-challenge2015.]% (rorre gniniart 20 10 00 1 2 3 4 5 6 interer.(1e4))% (rorre t set 56-layers 20-layers 56-layers 20-layer 图一.
<a id="S0016"></a> Source: p.1 S0016
Original: Training error (left) and test error (right) on CIFAR-10 with 20-layer and 56-layer “plain” networks.
中文: CIFAR-10上的训练出错(左)和测试出错(右),有20层和56层“平面”网络。
<a id="S0017"></a> Source: p.1 S0017
Original: The deeper network has higher training error, and thus test error.
中文: 更深的网络有更高的训练出错,因此测试出错.
<a id="S0018"></a> Source: p.1 S0018
Original: Similar phenomena on ImageNet is presented in Fig. 4. greatly benefited from very deep models.
中文: 图4介绍了ImageNet上的类似现象。
<a id="S0019"></a> Source: p.1 S0019
Original: Driven by the significance of depth, a question arises: Is learning better networks as easy as stacking more layers?
中文: 由于深度的重要性,出现了一个问题: 学习更好的网络像堆放更多的地层一样容易吗?
<a id="S0020"></a> Source: p.1 S0020
Original: An obstacle to answering this question was the notorious problem of vanishing/exploding gradients [1, 9], which hamper convergence from the beginning.
中文: 回答这个问题的一个障碍是臭名昭著的渐变[1、9]消失/爆炸问题,这从一开始就阻碍了趋同。
<a id="S0021"></a> Source: p.1 S0021
Original: This problem, however, has been largely addressed by normalized initialization [23, 9, 37, 13] and intermediate normalization layers [16], which enable networks with tens of layers to start converging for stochastic gradient descent (SGD) with backpropagation [22].
中文: 然而,这个问题在很大程度上通过正常化初始化[23,9,37,13]和中间正态层[16]来解决,这使得有数十层的网络能够开始以回向传播来趋同于分层梯度回落(SGD)[22].
<a id="S0022"></a> Source: p.1 S0022
Original: When deeper networks are able to start converging, a degradation problem has been exposed: with the network depth increasing, accuracy gets saturated (which might be unsurprising) and then degrades rapidly.
中文: 当更深层的网络能够开始趋同时,一个退化问题就暴露了:随着网络深度的增加,精度会饱和(这也许并不奇怪)并会迅速降解.
<a id="S0023"></a> Source: p.1 S0023
Original: Unexpectedly, such degradation is not caused by overfitting, and adding more layers to a suitably deep model leads to higher training error, as reported in [11, 42] and thoroughly verified by our experiments.
中文: 意外的是,这种退化不是由过度适应造成的,在适当深的模型中增加多层会导致更高的训练出错,如[11、42]所报告并经我们的实验充分核实。
<a id="S0024"></a> Source: p.1 S0024
Original: The degradation (of training accuracy) indicates that not all systems are similarly easy to optimize.
中文: (培训准确性的)退化表明,并非所有系统都同样容易优化。
<a id="S0025"></a> Source: p.1 S0025
Original: Let us consider a shallower architecture and its deeper counterpart that adds more layers onto it.
中文: 让我们考虑一个更浅的架构及其更深层的对应结构,使其增加更多的层次。
<a id="S0026"></a> Source: p.1 S0026
Original: There exists a solution by construction to the deeper model: the added layers are identity mapping, and the other layers are copied from the learned shallower model.
中文: 通过构建到更深的模型中存在一个解决方案:所添加的地层是身份映射,而其他地层则是从所学到的更浅的模型中复制出来的.
<a id="S0027"></a> Source: p.1 S0027
Original: The existence of this constructed solution indicates that a deeper model should produce no higher training error than its shallower counterpart.
中文: 这种已构建的解决方案的存在表明更深的模型不应产生比更浅的对应器更高的训练出错.
<a id="S0028"></a> Source: p.1 S0028
Original: But experiments show that our current solvers on hand are unable to find solutions that 1 5102 ceD 01 ]VC.sc[ 1v58330.2151:viXra
中文: 但实验显示,我们目前的解析器无法找到1 5102 ceD 01]VC.sc[1v58330.2151:viXra]的解决方案.
<a id="S0029"></a> Source: p.2 S0029
Original: ImageNet test set, and won the 1st place in the ILSVRC x 2015 classification competition.
中文: ImageNet测试集,并获得ILSVRCx2015分级赛第1名.
<a id="S0030"></a> Source: p.2 S0030
Original: The extremely deep repweight layer resentations also have excellent generalization performance F (x) relu x on other recognition tasks, and lead us to further win the weight layer identity 1st places on: ImageNet detection, ImageNet localization, COCO detection, and COCO segmentation in ILSVRC & (x)(cid:1)+(cid:1)x F relu COCO 2015 competitions.
中文: 极深的重分层怨恨在其它识别任务上也有出色的概括性能F (x) relu x,并带领我们进一步赢得了重分层身份的第1位:ImageNet检测,ImageNet本地化,COCO检测,以及COCO在ILSVRC & (x (cid:1)+ (cid:1)x F relu CO 2015比赛中的分层.
<a id="S0031"></a> Source: p.2 S0031
Original: This strong evidence shows that Figure 2.
中文: 这一有力证据表明图2。
<a id="S0032"></a> Source: p.2 S0032
Original: Residual learning: a building block. the residual learning principle is generic, and we expect that it is applicable in other vision and non-vision problems. are comparably good or better than the constructed solution (or unable to do so in feasible time). 2.
中文: 剩余学习:一个构件. 剩余学习原则是一般性的,我们期望它适用于其他愿景和非愿景问题。 (或无法在可行时间内这样做)。 2. 联合国
<a id="S0033"></a> Source: p.2 S0033
Original: Related Work In this paper, we address the degradation problem by introducing a deep residual learning framework.
中文: 相关工作 在本文中,我们通过引入深层残余学习框架来解决退化问题。
<a id="S0034"></a> Source: p.2 S0034
Original: In image recognition, VLAD stead of hoping each few stacked layers directly fit a [18] is a representation that encodes by the residual vectors desired underlying mapping, we explicitly let these lay- with respect to a dictionary, and Fisher Vector [30] can be ers fit a residual mapping.
中文: 在图像识别方面,VLAD(VLAD)代替希望每几个堆叠的地层直接适合一个[18]是一个由所希望的地基映射的残余向量编码的表示,我们明确允许这些与字典相连接,而Fisher Vector[30]可以是ers相接地基映射.
<a id="S0035"></a> Source: p.2 S0035
Original: Formally, denoting the desired formulated as a probabilistic version [18] of VLAD.
中文: 形式上,将所希望的表示成VLAD的概率版本[18].
<a id="S0036"></a> Source: p.2 S0036
Original: Both underlying mapping as (x), we let the stacked nonlinear of them are powerful shallow representations for image relayers fit another mappin H g of (x) := (x) x.
中文: 两种基础映射都作为(x),我们让堆叠的非线性它们成为图像中继器的强势浅层表示,适合(x)=(x)x的另一个映射H克.
<a id="S0037"></a> Source: p.2 S0037
Original: The orig- trieval and classification [4, 48].
中文: orig-trieval和分类[4,48].
<a id="S0038"></a> Source: p.2 S0038
Original: For vector quantization, inal mapping is recast into ( F x)+x. W H e hyp − othesize that it encoding residual vectors [17] is shown to be more effecis easier to optimize the re F sidual mapping than to optimize tive than encoding original vectors. the original, unreferenced mapping.
中文: 对于向量量化,内映射被重塑为(F x)+x. W H e hyp − othesize 表示它编码了剩余向量[17] , 显示比起编码原始向量,它更容易优化 re F sidual 映射. 原始的,无参考的绘图。
<a id="S0039"></a> Source: p.2 S0039
Original: To the extreme, if an In low-level vision and computer graphics, for solvidentity mapping were optimal, it would be easier to push ing Partial Differential Equations (PDEs), the widely used the residual to zero than to fit an identity mapping by a stack Multigrid method [3] reformulates the system as subprobof nonlinear layers. lems at multiple scales, where each subproblem is responsible for the residual solution between a coarser and a finer The formulation of (x) + x can be realized by feedfor- F scale.
中文: 至极,如果一个在低等视觉和计算机图形中,为了求解映射是最佳的,那么推出部分差异方程(PDEs)会更容易,被广泛使用的剩余到零,而不是用堆栈多网格法来搭配身份映射[3]将系统重塑为非线性分层. lems 在多尺度中,每个子问题负责一个相干器和细管之间的剩余溶液. (x) + x的配体可以通过feedfor-F尺度实现.
<a id="S0040"></a> Source: p.2 S0040
Original: An alternative to Multigrid is hierarchical basis preward neural networks with “shortcut connections” (Fig. 2). conditioning [45, 46], which relies on variables that repre- Shortcut connections [2, 34, 49] are those skipping one or sent residual vectors between two scales.
中文: Multigrid的一个替代办法是具有“短连接”的分级基础前神经网络(图2)。 条件[45,46],它依赖于可重排快捷连接[2,34,49]的变量,是跳过一个或发送到两个尺度之间的剩余向量.
<a id="S0041"></a> Source: p.2 S0041
Original: In our case, the shortcut connections simply [3, 45, 46] that these solvers converge much faster than stanperform identity mapping, and their outputs are added to dard solvers that are unaware of the residual nature of the the outputs of the stacked layers (Fig. 2).
中文: 在我们的情况中,这些解析器的快捷连接简单[3,45,46]的聚合速度远快于stanperform身份映射,其输出被添加到没有意识到堆叠层输出的剩余性质的达德解析器上(图2).
<a id="S0042"></a> Source: p.2 S0042
Original: These methods suggest that a good reformulation cut connections add neither extra parameter nor computaor preconditioning can simplify the optimization. tional complexity.
中文: 这些方法表明,良好的再造切开连接既不会增加额外的参数,也不会增加压缩前提,从而可以简化优化. 神经上的复杂性。
<a id="S0043"></a> Source: p.2 S0043
Original: The entire network can still be trained end-to-end by SGD with backpropagation, and can be eas- Shortcut Connections.
中文: 整个网络仍然可以由SGD通过回放来训练端到端,并且可以成为eas-快捷连接.
<a id="S0044"></a> Source: p.2 S0044
Original: Practices and theories that lead to ily implemented using common libraries (e.g., Caffe [19]) shortcut connections [2, 34, 49] have been studied for a long without modifying the solvers. time.
中文: 导致使用共同库(如Caffe[19])快捷连接[2,34,49]的ily执行的实践和理论,在不修改解析器的情况下被长期研究. 时间。
<a id="S0045"></a> Source: p.2 S0045
Original: An early practice of training multi-layer perceptrons We present comprehensive experiments on ImageNet (MLPs) is to add a linear layer connected from the network [36] to show the degradation problem and evaluate our input to the output [34, 49].
中文: 早期培训多层感官的做法 我们在ImageNet(MLPs)上做了全面的实验,即从网络上添加一个连接的线性层[36],以显示降解问题并评价我们对输出的投入[34,49].
<a id="S0046"></a> Source: p.2 S0046
Original: We show that: 1) Our extremely deep residual nets diate layers are directly connected to auxiliary classifiers are easy to optimize, but the counterpart “plain” nets (that for addressing vanishing/exploding gradients.
中文: 我们表明:(1) 我们极深的残余鱼网分层直接连接到辅助分类器上,很容易优化,但对应的 " 平地 " 网(用于处理已消失/正在消失的梯度)。
<a id="S0047"></a> Source: p.2 S0047
Original: The papers simply stack layers) exhibit higher training error when the of [39, 38, 31, 47] propose methods for centering layer redepth increases; 2) Our deep residual nets can easily enjoy sponses, gradients, and propagated errors, implemented by accuracy gains from greatly increased depth, producing re- shortcut connections.
中文: 当[39、38、31、47] 的论文提出更深层中枢的方法时,论文的堆叠层)显示出较高的训练错误; 2) 我们的深层残余网可以很容易地享受从深度大为增加的精度增加而得到的接通性、梯度和传播出的错误。
<a id="S0048"></a> Source: p.2 S0048
Original: In [44], an “inception” layer is comsults substantially better than previous networks. posed of a shortcut branch and a few deeper branches.
中文: 在[44]中,“概念”层比以前的网络要好得多。 由快捷分支和几根更深的分支构成。
<a id="S0049"></a> Source: p.2 S0049
Original: Similar phenomena are also shown on the CIFAR-10 set Concurrent with our work, “highway networks” [42, 43] [20], suggesting that the optimization difficulties and the present shortcut connections with gating functions [15]. effects of our method are not just akin to a particular dataset.
中文: 在CIFAR-10设定与我们的工作同时进行的“高速公路网络”[42,43] [20]中也显示了类似现象,这表明优化困难和目前捷径与地标函数的联系[15]. 我们的方法的效果 不仅仅是像一个特定的数据集。
<a id="S0050"></a> Source: p.2 S0050
Original: These gates are data-dependent and have parameters, in We present successfully trained models on this dataset with contrast to our identity shortcuts that are parameter-free. over 100 layers, and explore models with over 1000 layers.
中文: 这些闸门是数据依赖的,并有参数,我们在这个数据集上成功地展示了训练有素的模型,与我们的身份快捷键相对,这些快捷键是无参数的. 探索1000多层的模型
<a id="S0051"></a> Source: p.2 S0051
Original: When a gated shortcut is “closed” (approaching zero), the On the ImageNet classification dataset [36], we obtain layers in highway networks represent non-residual funcexcellent results by extremely deep residual nets.
中文: 当一个有门的捷径被“关闭”(接近零),即“On ImageNet”分类数据集[36]时,我们在高速公路网络中获得的分层代表着极其深厚的残留网产生的非活性真菌成果。
<a id="S0052"></a> Source: p.2 S0052
Original: On the contrary, our formulation always learns layer residual net is the deepest network ever presented on residual functions; our identity shortcuts are never closed, ImageNet, while still having lower complexity than VGG and all information is always passed through, with addinets [41].
中文: 相反,我们的表述总是学习层残网是有史以来最深的关于残余函数的网络;我们的身份快捷方式从未关闭过,ImageNet,而其复杂性仍然低于VGG,所有信息总是通过,并带有附加物[41].
<a id="S0053"></a> Source: p.2 S0053
Original: Our ensemble has 3.57% top-5 error on the tional residual functions to be learned.
中文: 我们的综艺节目在需要学习的电离残存功能上有3.57%的前5个出错.
<a id="S0054"></a> Source: p.3 S0054
Original: way networks have not demonstrated accuracy gains with ReLU [29] and the biases are omitted for simplifying noextremely increased depth (e.g., over 100 layers). tations.
中文: 网络在ReLU[29]中未显示准确性增益,为了简化新近增加的深度(如超过100层),略去偏差. t.
<a id="S0055"></a> Source: p.3 S0055
Original: The operation + x is performed by a shortcut F connection and element-wise addition.
中文: 操作+x由快捷键F连接和元素相加来完成.
<a id="S0056"></a> Source: p.3 S0056
Original: Deep Residual Learning ond nonlinearity after the addition (i.e., σ(y), see Fig. 2).
中文: 添加后的深残留学习非线性(即: " (y) " ,见图2)。
<a id="S0057"></a> Source: p.3 S0057
Original: The shortcut connections in Eqn.(1) introduce neither ex- 3.1.
中文: Eqn.(1)中的快捷连接既不引入前3.1.
<a id="S0058"></a> Source: p.3 S0058
Original: Residual Learning tra parameter nor computation complexity.
中文: 剩余学习参数或计算复杂性。
<a id="S0059"></a> Source: p.3 S0059
Original: This is not only Let us consider (x) as an underlying mapping to be attractive in practice but also important in our comparisons H fit by a few stacked layers (not necessarily the entire net), between plain and residual networks.
中文: 这不仅让我们认为(x)是一种在实践中具有吸引力的基础绘图,而且对我们用几层堆放(不一定是整个网)来比较平地网络和残余网络也很重要。
<a id="S0060"></a> Source: p.3 S0060
Original: We can fairly comwith x denoting the inputs to the first of these layers.
中文: 我们可以相当配合X 表示这些层中第一层的投入。
<a id="S0061"></a> Source: p.3 S0061
Original: If one pare plain/residual networks that simultaneously have the hypothesizes that multiple nonlinear layers can asymptoti- same number of parameters, depth, width, and computacally approximate complicated functions2, then it is equiv- tional cost (except for the negligible element-wise addition). alent to hypothesize that they can asymptotically approxi- The dimensions of x and must be equal in Eqn.(1). F mate the residual functions, i.e., (x) x (assuming that If this is not the case (e.g., when changing the input/output H − the input and output are of the same dimensions).
中文: 如果一个平原/剩余网络同时具有多个非线性地层可以同比地调出相同数量参数,深度,宽度,并比较近似复杂函数2的假说,那么它就是等效-通量成本(可忽略不计的元素加法除外). 完全可以假设他们可以无症状的近似... x的尺寸,必须在 Eqn.(1). F使剩余函数相配,即(x)x(假设情况并非如此(例如,在改变输入/输出H−输入和输出为同维度时)).
<a id="S0062"></a> Source: p.3 S0062
Original: So channels), we can perform a linear projection W s by the rather than expect stacked layers to approximate (x), we shortcut connections to match the dimensions: H explicitly let these layers approximate a residual function (x) := (x) x.
中文: 因此通道,我们可以通过线性投影Ws来进行,而不是期望堆叠的地层大致(x),我们快捷连接来匹配维度:H明确让这些地层接近一个剩余函数(x)=(x)x.
<a id="S0063"></a> Source: p.3 S0063
Original: The original function thus becomes y = (x, W i ) + W s x. (2) F H − F { } (x)+x.
中文: 原函数由此成为y= (x, Wi) + W s x (2) F H- F {} (x) + x.
<a id="S0064"></a> Source: p.3 S0064
Original: Although both forms should be able to asymptot- F We can also use a square matrix W in Eqn.(1).
中文: 虽然两种表型都应该能够同化-F,但我们也可以使用Eqn.(1)中的平方矩阵W.
<a id="S0065"></a> Source: p.3 S0065
Original: But we will ically approximate the desired functions (as hypothesized), s show by experiments that the identity mapping is sufficient the ease of learning might be different. for addressing the degradation problem and is economical, This reformulation is motivated by the counterintuitive and thus W is only used when matching dimensions. phenomena about the degradation problem (Fig. 1, left).
中文: 但是,我们将在理论上接近所期望的功能(假设的),通过实验表明,身份映射足够了,学习的方便程度可能有所不同。 为了解决退化问题,这种重构是因反直觉引起的,因此,W只在与维相匹配时才被使用。 关于退化问题的现象(图1,左)。
<a id="S0066"></a> Source: p.3 S0066
Original: As s The form of the residual function is flexible.
中文: 剩余函数的形式是灵活的。
<a id="S0067"></a> Source: p.3 S0067
Original: Experwe discussed in the introduction, if the added layers can F iments in this paper involve a function that has two or be constructed as identity mappings, a deeper model should F three layers (Fig. 5), while more layers are possible.
中文: Excerwe在导言中讨论过,如果本文中添加的地层可以F imments涉及一个具有两个功能或被构建为身份映射的功能,一个更深的模型应该有F 三地层(图5),而更多的地层是可能的.
<a id="S0068"></a> Source: p.3 S0068
Original: But if have training error no greater than its shallower counterhas only a single layer, Eqn.(1) is similar to a linear layer: part.
中文: 但如果训练有误不超过其更浅的反相只有一分层,则Eqn.(1)类似于一分线层:part.
<a id="S0069"></a> Source: p.3 S0069
Original: The degradation problem suggests that the solvers F y = W x + x, for which we have not observed advantages. might have difficulties in approximating identity mappings 1 We also note that although the above notations are about by multiple nonlinear layers.
中文: 降解问题表明解子F y = W x + x,对此我们没有观察到优势. 1 我们还注意到,虽然上述注释是由多层非线性图层构成的。
<a id="S0070"></a> Source: p.3 S0070
Original: With the residual learning refully-connected layers for simplicity, they are applicable to formulation, if identity mappings are optimal, the solvers convolutional layers.
中文: 将剩余学习重新连接的地层用于简单化,这些地层适用于配体,如果身份映射是最佳的,则解析者会分出地层.
<a id="S0071"></a> Source: p.3 S0071
Original: The function (x, W ) can repremay simply drive the weights of the multiple nonlinear lay- i F { } sent multiple convolutional layers.
中文: 函数 (x, W) 可以直接驱动多个非线性平地- i F { } 发送多个分层的重心.
<a id="S0072"></a> Source: p.3 S0072
Original: The element-wise addiers toward zero to approach identity mappings. tion is performed on two feature maps, channel by channel.
中文: 元素辅助器向零方向接近身份映射。 tv在两个专题地图上进行,按频道进行.
<a id="S0073"></a> Source: p.3 S0073
Original: In real cases, it is unlikely that identity mappings are optimal, but our reformulation may help to precondition the 3.3.
中文: 在实际情况下,身份测绘不可能是最佳的,但我们的重新表述可能有助于为3.3设定先决条件。
<a id="S0074"></a> Source: p.3 S0074
Original: If the optimal function is closer to an identity mapping than to a zero mapping, it should be easier for the We have tested various plain/residual nets, and have obsolver to find the perturbations with reference to an identity served consistent phenomena.
中文: 如果最佳功能更接近于身份映射,而不是零映射,那么对于我们已经测试了各种平地/剩余网,并且对于寻找身份服务于一致现象的扰动就比较容易。
<a id="S0075"></a> Source: p.3 S0075
Original: To provide instances for dismapping, than to learn the function as a new one.
中文: 提供拆分实例,而不是学习新的功能。
<a id="S0076"></a> Source: p.3 S0076
Original: We show cussion, we describe two models for ImageNet as follows. by experiments (Fig. 7) that the learned residual functions in Plain Network.
中文: 我们显示打击,我们描述图像网的两个模型如下. (图7),即Plain Network中学到的剩余功能。
<a id="S0077"></a> Source: p.3 S0077
Original: Our plain baselines (Fig. 3, middle) are general have small responses, suggesting that identity mapmainly inspired by the philosophy of VGG nets [41] (Fig. 3, pings provide reasonable preconditioning. left).
中文: 我们的平地基线(图3,中地)一般有小的响应,说明身份图主要受VGG网哲学启发[41](图3,平地提供合理的前提条件. 页:1
<a id="S0078"></a> Source: p.3 S0078
Original: The convolutional layers mostly have 3 3 filters and × 3.2.
中文: 演化地层多有3个3个滤波器和× 3.2.
<a id="S0079"></a> Source: p.3 S0079
Original: Identity Mapping by Shortcuts follow two simple design rules: (i) for the same output feature map size, the layers have the same number of fil- We adopt residual learning to every few stacked layers. ters; and (ii) if the feature map size is halved, the num- A building block is shown in Fig. 2.
中文: 通过快捷键进行身份映射遵循了两个简单的设计规则:(一) 对于相同的输出特征映射大小,层有相同的fil-数量. 我们把残存的学习 推广到每一个堆积层。 之三;和 (二) 如果将特征图大小减半,则数字 - 图2显示了一个构件。
<a id="S0080"></a> Source: p.3 S0080
Original: Formally, in this paper ber of filters is doubled so as to preserve the time comwe consider a building block defined as: plexity per layer.
中文: 正式地说,在本文中,过滤器的泊位加倍,以便保留我们考虑一个构件的时间,该构件的定义是:每层复杂度.
<a id="S0081"></a> Source: p.3 S0081
Original: We perform downsampling directly by convolutional layers that have a stride of 2.
中文: 我们直接通过有2个阶地的分层来进行下采样。
<a id="S0082"></a> Source: p.3 S0082
Original: The network y = (x, W ) + x. (1) F { i } ends with a global average pooling layer and a 1000-way fully-connected layer with softmax.
中文: 网络 y = (x, W ) + x. (1) F {i } 以全球平均集合层和1000道全相接层相接而成,并带有软马克斯.
<a id="S0083"></a> Source: p.3 S0083
Original: The total number of Here x and y are the input and output vectors of the layweighted layers is 34 in Fig. 3 (middle). ers considered.
中文: Here x 和 y 为自重层的输入和输出向量的总数在图3(中)为34. 考虑过。
<a id="S0084"></a> Source: p.3 S0084
Original: The function (x, W ) represents the F { i } It is worth noticing that our model has fewer filters and residual mapping to be learned.
中文: 函数 (x, W) 代表 F { i } 值得注意的是,我们的模型的过滤器和可学习的剩余绘图较少。
<a id="S0085"></a> Source: p.3 S0085
Original: For the example in Fig. 2 lower complexity than VGG nets [41] (Fig. 3, left).
中文: 例如,图2中的复杂性低于VGG网41。
<a id="S0086"></a> Source: p.3 S0086
Original: Our 34that has two layers, = W σ(W x) in which σ denotes 2 1 F layer baseline has 3.6 billion FLOPs (multiply-adds), which 2This hypothesis, however, is still an open question.
中文: 我们的34个有两层,=W / (W x),其中以2 1 F 的分层基线表示有36亿FLOP(乘以-添加),然而,这一假设仍然是一个未决问题。
<a id="S0087"></a> Source: p.3 S0087
Original: See [28]. is only 18% of VGG-19 (19.6 billion FLOPs). 3
中文: 见[28]。 只有18%的VGG-19(196亿FLOP)。 3个
<a id="S0088"></a> Source: p.4 S0088
Original: VGG-19 34-layer plain 34-layer residual Residual Network.
中文: VGG-19型34层平原34层残余物网络.
<a id="S0089"></a> Source: p.4 S0089
Original: Based on the above plain network, we insert shortcut connections (Fig. 3, right) which turn the image image image s o iz u e t : p 2 u 2 t 4 3x3 conv, 64 network into its counterpart residual version.
中文: 基于上述平面网络,我们插入了快捷连接(图3,右),将图像图像 s o iz u e t: p 2 u 2 t 4 3x3 conv,64 network 转换成对应的剩余版本.
<a id="S0090"></a> Source: p.4 S0090
Original: The identity shortcuts (Eqn.(1)) can be directly used when the input and 3x3 conv, 64 output are of the same dimensions (solid line shortcuts in pool, /2 output Fig. 3).
中文: 身份快捷键(Eqn.(1))可以在输入和3x3凸起时直接被使用,64个输出的尺寸相同(集合中的固行快捷键,2输出图.
<a id="S0091"></a> Source: p.4 S0091
Original: When the dimensions increase (dotted line shortcuts size: 112 3x3 conv, 128 in Fig. 3), we consider two options: (A) The shortcut still 3x3 conv, 128 7x7 conv, 64, /2 7x7 conv, 64, /2 performs identity mapping, with extra zero entries padded pool, /2 pool, /2 pool, /2 output for increasing dimensions.
中文: 当维度增加(被点缀的行跑快道:112 3x3 conv,128 in Fig. 3)时,我们考虑两个选项: (A)快捷道仍然为3x3 conv,128 7x7 conv,64,/2 7x7 conv,64,2执行身份映射,额外零条目加码池,2池,2池,2输出来增加维度.
<a id="S0092"></a> Source: p.4 S0092
Original: This option introduces no extra size: 56 3x3 conv, 256 3x3 conv, 64 3x3 conv, 64 parameter; (B) The projection shortcut in Eqn.(2) is used to 3x3 conv, 256 3x3 conv, 64 3x3 conv, 64 match dimensions (done by 1 1 convolutions).
中文: 此选项不引入额外大小:56个3x3 conv,256个3x3 conv,64个3x3 conv,64个参数; (B) Eqn.(2)中的投影快捷键用于3x3 conv,256个3x3 conv,64个3x3 conv,64个相配尺寸(由 1 1 1 个进取).
<a id="S0093"></a> Source: p.4 S0093
Original: For both × 3x3 conv, 256 3x3 conv, 64 3x3 conv, 64 options, when the shortcuts go across feature maps of two 3x3 conv, 256 3x3 conv, 64 3x3 conv, 64 sizes, they are performed with a stride of 2. 3x3 conv, 64 3x3 conv, 64 3.4.
中文: 对于双相×3×3相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相
<a id="S0094"></a> Source: p.4 S0094
Original: Implementation 3x3 conv, 64 3x3 conv, 64 pool, /2 3x3 conv, 128, /2 3x3 conv, 128, /2 Our implementation for ImageNet follows the practice output size: 28 3x3 conv, 512 3x3 conv, 128 3x3 conv, 128 in [21, 41].
中文: 执行3×3收缩口,64个3×3收缩口,64个池,/2个3×3收缩口,128个,/2个3×3收缩口,128个,2个. 我们对ImageNet的执行遵循惯例输出大小:28个3×3收缩口,512个3×3收缩口,128个3×3收缩口,128个在[21,41].
<a id="S0095"></a> Source: p.4 S0095
Original: The image is resized with its shorter side ran- 3x3 conv, 512 3x3 conv, 128 3x3 conv, 128 domly sampled in [256, 480] for scale augmentation [41]. 3x3 conv, 512 3x3 conv, 128 3x3 conv, 128 A 224 224 crop is randomly sampled from an image or its × horizontal flip, with the per-pixel mean subtracted [21].
中文: 图像以更短的侧跑-3x3凸起,512个3x3凸起,128个3x3凸起,128个多米取样于[256,480]进行规模增强[41]. 3x3收缩口,512个3x3收缩口,128个3x3收缩口,128个A 224个收缩口,从一幅图像或其×平面翻接处随机抽取出,以每像素平均值减去[21].
<a id="S0096"></a> Source: p.4 S0096
Original: The 3x3 conv, 512 3x3 conv, 128 3x3 conv, 128 standard color augmentation in [21] is used.
中文: 3x3收缩口,512个3x3收缩口,128个3x3收缩口,128个标准色增收在[21]中使用.
<a id="S0097"></a> Source: p.4 S0097
Original: We adopt batch 3x3 conv, 128 3x3 conv, 128 normalization (BN) [16] right after each convolution and 3x3 conv, 128 3x3 conv, 128 before activation, following [16].
中文: 我们采用批量的3x3收缩,128个3x3收缩,128个正常化(BN)[16]在每次分化之后和3x3收缩,128个3x3收缩,128个活化前,16个活化后.
<a id="S0098"></a> Source: p.4 S0098
Original: We initialize the weights 3x3 conv, 128 3x3 conv, 128 as in [13] and train all plain/residual nets from scratch.
中文: 我们将重心3x3凸起,128 3x3凸起,128 如 [13] 从头开始训练所有平地/剩余网.
<a id="S0099"></a> Source: p.4 S0099
Original: We s o iz u e t : p 1 u 4 t pool, /2 3x3 conv, 256, /2 3x3 conv, 256, /2 use SGD with a mini-batch size of 256.
中文: 我们s o iz u e t: p 1 u 4 t pool, p 1 u 4 t pool, s/2 3x3 conv, 256, s/2 3x3 conv, 256, s 2. 使用 SGD, 微型批量大小为256.
<a id="S0100"></a> Source: p.4 S0100
Original: The learning rate 3x3 conv, 512 3x3 conv, 256 3x3 conv, 256 starts from 0.1 and is divided by 10 when the error plateaus, 3x3 conv, 512 3x3 conv, 256 3x3 conv, 256 and the models are trained for up to 60 104 iterations.
中文: 学习速率为:3x3 conv,512 3x3 conv,256 3x3 conv,256起自0.1起,在错误高原,3x3 conv,512 3x3 conv,256 3x3 conv,256和模型被训练最多60 104个迭代时被除去.
<a id="S0101"></a> Source: p.4 S0101
Original: We 3x3 conv, 512 3x3 conv, 256 3x3 conv, 256 × use a weight decay of 0.0001 and a momentum of 0.9.
中文: We 3x3 conv, 512 3x3 conv, 256 3x3 conv, 256 ×使用0.0001的重量衰变和0.9的动力.
<a id="S0102"></a> Source: p.4 S0102
Original: We 3x3 conv, 512 3x3 conv, 256 3x3 conv, 256 do not use dropout [14], following the practice in [16]. 3x3 conv, 256 3x3 conv, 256 In testing, for comparison studies we adopt the standard 3x3 conv, 256 3x3 conv, 256 10-crop testing [21].
中文: We 3x3 conv, 512 3x3 conv, 256 3x3 conv, 256 不使用退出 [14],遵循了 [16] 的惯例. 3x3 conv, 256 3x3 conv, 256 在测试中,为了比较研究,我们采用了标准3x3收缩,256个3x3收缩,256个10个收缩测试[21].
<a id="S0103"></a> Source: p.4 S0103
Original: For best results, we adopt the fully- 3x3 conv, 256 3x3 conv, 256 convolutional form as in [41, 13], and average the scores 3x3 conv, 256 3x3 conv, 256 at multiple scales (images are resized such that the shorter 3x3 conv, 256 3x3 conv, 256 side is in 224, 256, 384, 480, 640 ). { } 3x3 conv, 256 3x3 conv, 256 4.
中文: 为了取得最佳效果,我们采用[41,13]中的全-3x3 conv,256 3x3 conv,256等同分数,平均分数为:3x3 conv,256 3x3 conv,256等同分数分数分数分数分数分数分数分数分数分数取数分数分数分数分数分数分数分数分数分数分数分数分数分数取数分数分数分数分数分数分数分数取数分数分数分数分数分数分数取数分数分数分数分数分数分数分数分数分数取数分数分数分数分数分数分数分数分数分数分数分数分数分数分数取数分数分数分数分数分数分数分数分数分数分数分数分数分数分数分数分数分数分数分数分数分数分数分数分数分数分数分数分数分数分数分数分数分
<a id="S0104"></a> Source: p.4 S0104
Original: Experiments 3x3 conv, 256 3x3 conv, 256 o si u z t e p : u 7 t pool, /2 3x3 conv, 512, /2 3x3 conv, 512, /2 4.1.
中文: 实验 3x3 conv, 256 3x3 conv, 256 o si u z t e p: u 7 t pool, /2 3x3 conv, 512, /2 3x3 conv, 512, 2. 4.1.
<a id="S0105"></a> Source: p.4 S0105
Original: ImageNet Classification 3x3 conv, 512 3x3 conv, 512 We evaluate our method on the ImageNet 2012 classifi- 3x3 conv, 512 3x3 conv, 512 cation dataset [36] that consists of 1000 classes.
中文: ImageNet 分类 3x3 conv, 512 3x3 conv, 512 我们在2012 Classifi-3x3 conv, 512 3x3 conv, 512 cation数据集[36]上评价我们的方法,这些数据集由1000个类组成.
<a id="S0106"></a> Source: p.4 S0106
Original: The models 3x3 conv, 512 3x3 conv, 512 are trained on the 1.28 million training images, and evalu- 3x3 conv, 512 3x3 conv, 512 ated on the 50k validation images.
中文: 型号为3x3相接相,512相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相
<a id="S0107"></a> Source: p.4 S0107
Original: We also obtain a final 3x3 conv, 512 3x3 conv, 512 result on the 100k test images, reported by the test server. o si u z t e p : u 1 t fc 4096 avg pool avg pool We evaluate both top-1 and top-5 error rates. fc 4096 fc 1000 fc 1000 fc 1000 Plain Networks.
中文: 我们还获得一个最终的3x3凸起,512个3x3凸起,在测试服务器报告的100k测试图像上得到512个结果. o si u z t e p: u 1 t fc 4096 avg pool avg pool 我们评价上一和上五的错误率. fc 4096 fc 1000 fc 1000 fc 1000平原网络.
<a id="S0108"></a> Source: p.4 S0108
Original: We first evaluate 18-layer and 34-layer plain nets.
中文: 我们首先评估18层和34层平地网。
<a id="S0109"></a> Source: p.4 S0109
Original: The 34-layer plain net is in Fig. 3 (middle).
中文: 34层平原网为图3(中).
<a id="S0110"></a> Source: p.4 S0110
Original: Example network architectures for ImageNet.
中文: 图像网络的示例网络架构.
<a id="S0111"></a> Source: p.4 S0111
Original: Left: the 18-layer plain net is of a similar form.
中文: 左:18层平原网形式相近.
<a id="S0112"></a> Source: p.4 S0112
Original: See Table 1 for de- VGG-19 model [41] (19.6 billion FLOPs) as a reference.
中文: 参见表1中的去VGG-19型 [41] (196亿FLOPs)作为参考.
<a id="S0113"></a> Source: p.4 S0113
Original: Mid- tailed architectures. dle: a plain network with 34 parameter layers (3.6 billion FLOPs).
中文: 中尾建筑. dle:有34个参数层(36亿FLOP)的平地网络.
<a id="S0114"></a> Source: p.4 S0114
Original: The results in Table 2 show that the deeper 34-layer plain Right: a residual network with 34 parameter layers (3.6 billion net has higher validation error than the shallower 18-layer FLOPs).
中文: 表2的结果显示更深的34层平原右:一个有34个参数层的剩余网络(36亿网有比更浅的18层FLOPs更高的验证错误).
<a id="S0115"></a> Source: p.4 S0115
Original: The dotted shortcuts increase dimensions.
中文: 被点的快捷键会增加维度.
<a id="S0116"></a> Source: p.4 S0116
Original: To reveal the reasons, in Fig. 4 (left) we commore details and other variants. pare their training/validation errors during the training procedure.
中文: 为了揭示原因,在图4(左)中,我们比较详细和其他变体。 在培训过程中纠正他们的训练/鉴定错误。
<a id="S0117"></a> Source: p.4 S0117
Original: We have observed the degradation problem - the 4
中文: 我们观察到了退化问题 -- -- 4
<a id="S0118"></a> Source: p.5 S0118
Original: layer name output size 18-layer 34-layer 50-layer 101-layer 152-layer conv1 112 112 7 7, 64, stride 2 × × 3 3 max pool, stride 2 conv2 x 56 × 56 (cid:20) 3 3 × 3 3 , , 6 6 4 4 (cid:21) × 2 (cid:20) 3 3 × 3 3 , , 6 6 4 4 (cid:21) × 3 × 1 3 × × 1 3 , , 6 6 4 4 × 3 1 3 × × 1 3 , , 6 6 4 4 × 3 1 3 × × 1 3 , , 6 6 4 4 × 3 × × 1 1, 256 1 1, 256 1 1, 256 conv3 x 28 × 28 (cid:20) 3 3 × 3 3 , , 1 1 2 2 8 8 (cid:21) × 2 (cid:20) 3 3 × 3 3 , , 1 1 2 2 8 8 (cid:21) × 4 1 3 × × × 1 3 , , 1 1 2 2 8 8 × 4 1 3 × × × 1 3 , , 1 1 2 2 8 8 × 4 1 3 × × × 1 3 , , 1 1 2 2 8 8 × 8 × × 1 1, 512 1 1, 512 1 1, 512 conv4 x 14 × 14 (cid:20) 3 3 × 3 3 , , 2 2 5 5 6 6 (cid:21) × 2 (cid:20) 3 3 × 3 3 , , 2 2 5 5 6 6 (cid:21) × 6 1 3 × × × 1 3 , , 2 2 5 5 6 6 × 6 1 3 × × × 1 3 , , 2 2 5 5 6 6 × 23 1 3 × × × 1 3 , , 2 2 5 5 6 6 × 36 × × 1 1, 1024 1 1, 1024 1 1, 1024 conv5 x 7 × 7 (cid:20) 3 3 × 3 3 , , 5 5 1 1 2 2 (cid:21) × 2 (cid:20) 3 3 × 3 3 , , 5 5 1 1 2 2 (cid:21) × 3 1 3 × × × 1 3 , , 5 5 1 1 2 2 × 3 1 3 × × × 1 3 , , 5 5 1 1 2 2 × 3 1 3 × × × 1 3 , , 5 5 1 1 2 2 × 3 × × 1 1, 2048 1 1, 2048 1 1, 2048 × × × 1 1 average pool, 1000-d fc, softmax × FLOPs 1.8 109 3.6 109 3.8 109 7.6 109 11.3 109 × × × × × Table 1.
中文: 层名输出大小为18相 34相 50相 50相 相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相 , 1024 1 1, 1024 conv5 x 7 × 7 (Cid:20) 3 × 3 × 3 , 5 5 1 1 2 (cid:21) × 2 (cid:20) 3 × 3 3 3 · 3 , 5 5 1 1 1 2 (cid:21) × 3 · 3 · 3 · 3 · 3 · 3 · · 5 · 5 · 3 · · 3 · · 3 · · 1 · 2 · 3 · · 5 · · 5 · 3 · · · · 3 · 3 · · · · 3 · · · · 3 · · · · · · · · · · · · · · · · · · · · · · · · · ·
<a id="S0119"></a> Source: p.5 S0119
Original: Building blocks are shown in brackets (see also Fig. 5), with the numbers of blocks stacked.
中文: 块块以括号显示(又见图5),块块数被堆放.
<a id="S0120"></a> Source: p.5 S0120
Original: Downsampling is performed by conv3 1, conv4 1, and conv5 1 with a stride of 2. 60 50 40 30 20 0 10 20 30 40 50 iter. (1e4) )%( rorre 60 50 40 30 plain-18 plain-34 20 0 10 20 30 40 50 iter. (1e4) )%( rorre 34-layer 18-layer 18-layer ResNet-18 ResNet-34 34-layer Figure 4.
中文: 俯冲采样由凸起相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接相接
<a id="S0121"></a> Source: p.5 S0121
Original: Thin curves denote training error, and bold curves denote validation error of the center crops.
中文: 细曲线指训练出错,而粗曲线指中心收成的验证出错.
<a id="S0122"></a> Source: p.5 S0122
Original: Left: plain networks of 18 and 34 layers.
中文: 左:由18层和34层组成的平地网.
<a id="S0123"></a> Source: p.5 S0123
Original: In this plot, the residual networks have no extra parameter compared to their plain counterparts. plain ResNet reducing of the training error3.
中文: 在这个情节中,相较于平庸的对应物,剩余网络没有额外的参数. plain ResNet 减少训练错误 3。
<a id="S0124"></a> Source: p.5 S0124
Original: The reason for such opti- 18 layers 27.94 27.88 mization difficulties will be studied in the future. 34 layers 28.54 25.03 Residual Networks.
中文: 今后将研究造成这种Opi-18层27.94 27.88米化困难的原因。 34层 28.54 25.03 残存网.
<a id="S0125"></a> Source: p.5 S0125
Original: Next we evaluate 18-layer and 34- Table 2.
中文: 接下来我们评价18层和34层表2。
<a id="S0126"></a> Source: p.5 S0126
Original: Top-1 error (%, 10-crop testing) on ImageNet validation. layer residual nets (ResNets).
中文: 在ImageNet验证上,Top-1出错(%,10crop测试). 层余网。
<a id="S0127"></a> Source: p.5 S0127
Original: The baseline architectures Here the ResNets have no extra parameter compared to their plain are the same as the above plain nets, expect that a shortcut counterparts.
中文: 基线结构 这里ResNet与平地相比没有额外的参数,与上面的平地网是相同的,希望有一个快捷方式对应的.
<a id="S0128"></a> Source: p.5 S0128
Original: Fig. 4 shows the training procedures. connection is added to each pair of 3 3 filters as in Fig. 3 × (right).
中文: 图4显示了培训程序。 连接被添加到每对3 3 个过滤器中, 如图 3 × (右) 。
<a id="S0129"></a> Source: p.5 S0129
Original: In the first comparison (Table 2 and Fig. 4 right), we use identity mapping for all shortcuts and zero-padding 34-layer plain net has higher training error throughout the for increasing dimensions (option A).
中文: 在第一次比较(表2和图4右侧)中,我们对所有捷径使用身份映射,零平面34层平面网在增加维度的整个过程中有较高的训练错误(选项A)。
<a id="S0130"></a> Source: p.5 S0130
Original: So they have no extra whole training procedure, even though the solution space parameter compared to the plain counterparts. of the 18-layer plain network is a subspace of that of the We have three major observations from Table 2 and 34-layer one.
中文: 因此他们没有额外的整个训练程序,即使与平地对应方相比的溶液空间参数. 18层平原网络是一个子空间,我们从表2和表34层的1中得出了三大观测结果。
<a id="S0131"></a> Source: p.5 S0131
Original: First, the situation is reversed with residual learn- We argue that this optimization difficulty is unlikely to ing – the 34-layer ResNet is better than the 18-layer ResNet be caused by vanishing gradients.
中文: 第一,情况被扭转 剩下的学习 - 我们主张这种优化难度不太可能发生 — — 34层的ResNet比18层的ResNet更好,
<a id="S0132"></a> Source: p.5 S0132
Original: More importantly, the 34-layer ResNet exhibits trained with BN [16], which ensures forward propagated considerably lower training error and is generalizable to the signals to have non-zero variances.
中文: 更重要的是,通过BN[16]培训的34层ResNet展品,确保了远期传播的训练出更低的训练出错,并且对信号普遍来说有非零差异。
<a id="S0133"></a> Source: p.5 S0133
Original: We also verify that the validation data.
中文: 我们还核实验证数据。
<a id="S0134"></a> Source: p.5 S0134
Original: This indicates that the degradation problem backward propagated gradients exhibit healthy norms with is well addressed in this setting and we manage to obtain BN.
中文: 这表明,退化问题向后扩散的梯度呈现出健康的规范,在这种背景下得到了很好的处理,我们设法获得了生物网络。
<a id="S0135"></a> Source: p.5 S0135
Original: So neither forward nor backward signals vanish.
中文: 因此,前向信号和后向信号都没有消失。
<a id="S0136"></a> Source: p.5 S0136
Original: In accuracy gains from increased depth. fact, the 34-layer plain net is still able to achieve compet- Second, compared to its plain counterpart, the 34-layer itive accuracy (Table 3), suggesting that the solver works 3We have experimented with more training iterations (3×) and still obto some extent.
中文: 从深度增加中获得准确性。 事实上,34层平原网 仍然能够实现竞争... 第二,与其平面对应的34层活性精度(表3)相比,说明"解决者"的作品"3"(We)已经尝试了更多的训练活性迭代(3×),并在某种程度上仍然模糊.
<a id="S0137"></a> Source: p.5 S0137
Original: We conjecture that the deep plain nets may served the degradation problem, suggesting that this problem cannot be have exponentially low convergence rates, which impact the feasibly addressed by simply using more iterations. 5
中文: 我们推测,深平原网可能对退化问题有利,这表明这一问题不可能有指数低的汇合率,仅仅使用更多的迭代就会影响实际解决的问题。 页:1
<a id="S0138"></a> Source: p.6 S0138
Original: VGG-16 [41] 28.07 9.33 3x3, 64 1x1, 64 relu GoogLeNet [44] - 9.15 relu 3x3, 64 PReLU-net [13] 24.27 7.38 3x3, 64 relu 1x1, 256 plain-34 28.54 10.02 relu relu ResNet-34 A 25.03 7.76 ResNet-34 B 24.52 7.46 Figure 5. A deeper residual function F for ImageNet.
中文: VGG-16 [41] 28.07 9.33 3x3, 64 1x1, 64 relu GoogleNet [44] - 9.15 relu 3x3, 64 PRELU-net [13] 24.27 7.38 3x3, 64 relu 1x1, 256 plain-34 28.54 10.02 relu relu relu ResNet-34 A 25.03 7.76 ResNet-34 B 24.52 7.46 图5. 图像网更深的剩余功能 F.
<a id="S0139"></a> Source: p.6 S0139
Original: Left: a ResNet-34 C 24.19 7.40 building block (on 56×56 feature maps) as in Fig. 3 for ResNet- ResNet-50 22.85 6.71 34.
中文: 左:一个ResNet-34 C 24.19 7.40 构件(在56×56地物图上),如图3中的ResNet-ResNet-50 22.85 6.71 34.
<a id="S0140"></a> Source: p.6 S0140
Original: Right: a “bottleneck” building block for ResNet-50/101/152.
中文: 右:ResNet-50/101/152的“瓶颈”构件。
<a id="S0141"></a> Source: p.6 S0141
Original: ResNet-101 21.75 6.05 ResNet-152 21.43 5.71 parameter-free, identity shortcuts help with training.
中文: ResNet-101 21.75 6.05 ResNet-152 21.43 5.71 无参数,身份快捷键帮助培训.
<a id="S0142"></a> Source: p.6 S0142
Original: Error rates (%, 10-crop testing) on ImageNet validation. we investigate projection shortcuts (Eqn.(2)).
中文: ImageNet验证错误率 (%, 10 crop testing). 我们调查投影捷径(Eqn.(2))。
<a id="S0143"></a> Source: p.6 S0143
Original: In Table 3 we VGG-16 is based on our test.
中文: 在表3中,我们的VGG-16是基于我们的测试。
<a id="S0144"></a> Source: p.6 S0144
Original: ResNet-50/101/152 are of option B compare three options: (A) zero-padding shortcuts are used that only uses projections for increasing dimensions. for increasing dimensions, and all shortcuts are parameterfree (the same as Table 2 and Fig. 4 right); (B) projecmethod top-1 err. top-5 err.
中文: ResNet-50/101/152是备选方案B的比较方案3:(A) 使用零打字捷径,只使用预测来增加尺寸。 用于增加维度,所有快捷键都是无参数的(与表2和图4右同); (B) projecmod top-1 错误; top-5 错误。
<a id="S0145"></a> Source: p.6 S0145
Original: VGG [41] (ILSVRC’14) - 8.43† tion shortcuts are used for increasing dimensions, and other shortcuts are identity; and (C) all shortcuts are projections.
中文: VGG [41] (ILSVRC'14) - 8.43†通取快捷键用于增加维度,而其他快捷键是身份;和 (C) 所有快捷键都是预测.
<a id="S0146"></a> Source: p.6 S0146
Original: GoogLeNet [44] (ILSVRC’14) - 7.89 Table 3 shows that all three options are considerably bet- VGG [41] (v5) 24.4 7.1 ter than the plain counterpart. B is slightly better than A.
中文: GoogleNet [44] (ILSVRC ' 14) - 7.89 表3显示,所有三种选择都比普通对应方案下注相当多 -- -- VGG[41] (v5) 24.4 7.1之三。 B稍好于A.
<a id="S0147"></a> Source: p.6 S0147
Original: We PReLU-net [13] 21.59 5.71 argue that this is because the zero-padded dimensions in A BN-inception [16] 21.99 5.81 indeed have no residual learning. C is marginally better than ResNet-34 B 21.84 5.71 B, and we attribute this to the extra parameters introduced ResNet-34 C 21.53 5.60 by many (thirteen) projection shortcuts.
中文: 我们PRELU-net [13] 21.59 5.71认为,这是因为A BN-inception [16] 21.99 5.81中的零加分维数确实没有剩余学习. C略好于ResNet-34 B 21.84 5.71 B,我们将此归因于许多(十三个)投影捷径引入ResNet-34 C 21.53 5.60的额外参数.
<a id="S0148"></a> Source: p.6 S0148
Original: But the small dif- ResNet-50 20.74 5.25 ferences among A/B/C indicate that projection shortcuts are ResNet-101 19.87 4.60 not essential for addressing the degradation problem.
中文: 但是,A/B/C之间的小差-ResNet-50 20.74 5.25推论表明,投影捷径是ResNet-101 19.87 4.60,对于解决退化问题并不重要。
<a id="S0149"></a> Source: p.6 S0149
Original: So we ResNet-152 19.38 4.49 do not use option C in the rest of this paper, to reduce mem- Table 4.
中文: 因此,我们ResNet-152 19.38 4.49在本文其余部分不使用备选案文C来减少迷你表4。
<a id="S0150"></a> Source: p.6 S0150
Original: Error rates (%) of single-model results on the ImageNet ory/time complexity and model sizes.
中文: ImageNet ory/时间复杂度和模型大小上的单模型结果出错率 (%).
<a id="S0151"></a> Source: p.6 S0151
Original: Identity shortcuts are validation set (except † reported on the test set). particularly important for not increasing the complexity of the bottleneck architectures that are introduced below. method top-5 err. (test) VGG [41] (ILSVRC’14) 7.32 Deeper Bottleneck Architectures.
中文: 身份快捷键是验证集(测试集上报告的除外)。 对于不增加下文提出的瓶颈结构的复杂性特别重要。 (测试) VGG [41] (ILSVRC ' 14) 7.32 Deeper Bottleneck Architectures.
<a id="S0152"></a> Source: p.6 S0152
Original: Next we describe our GoogLeNet [44] (ILSVRC’14) 6.66 deeper nets for ImageNet.
中文: 我们接下来描述我们的GoogleNet446.66个更深的ImageNet网。
<a id="S0153"></a> Source: p.6 S0153
Original: Because of concerns on the train- VGG [41] (v5) 6.8 ing time that we can afford, we modify the building block PReLU-net [13] 4.94 as a bottleneck design4.
中文: 由于对列车-VGG41 6.8发车时间的担忧,我们修改了PRELU-net[13] 4.94号构件作为瓶颈设计4.
<a id="S0154"></a> Source: p.6 S0154
Original: For each residual function , we F BN-inception [16] 4.82 use a stack of 3 layers instead of 2 (Fig. 5).
中文: 对于每个剩余函数,我们F BN-inception [16] 4.82使用一叠由3层而不是2层(图5.
<a id="S0155"></a> Source: p.6 S0155
Original: The three layers ResNet (ILSVRC’15) 3.57 are 1 1, 3 3, and 1 1 convolutions, where the 1 1 layers × × × × are responsible for reducing and then increasing (restoring) Table 5.
中文: 三层ResNet (ILSVRC ' 15) 3.57是 1 1, 3 3和 1 1 的演化,其中 1 1 个层 × × × 负责减少并再增加(恢复) 表 5.
<a id="S0156"></a> Source: p.6 S0156
Original: The top-5 error is on the dimensions, leaving the 3 3 layer a bottleneck with smaller test set of ImageNet and reported by the test server. × input/output dimensions.
中文: 上出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出入出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出入出出出出出出出出出出出出出出出出出入出出出出出出出出出出出出出出出出出出出出出出出出出入出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出出 × 投入/产出维度。
<a id="S0157"></a> Source: p.6 S0157
Original: Fig. 5 shows an example, where both designs have similar time complexity.
中文: 图5显示了一个实例,两种设计的时间复杂度相近.
<a id="S0158"></a> Source: p.6 S0158
Original: ResNet reduces the top-1 error by 3.5% (Table 2), resulting The parameter-free identity shortcuts are particularly imfrom the successfully reduced training error (Fig. 4 right vs. portant for the bottleneck architectures.
中文: ResNet将前1个错误减少3.5%(表2),结果 无参数身份快捷键尤其出自成功减少的训练出错(图4右对接瓶颈架构的出错).
<a id="S0159"></a> Source: p.6 S0159
Original: This comparison verifies the effectiveness of residual cut in Fig. 5 (right) is replaced with projection, one can learning on extremely deep systems. show that the time complexity and model size are doubled, as the shortcut is connected to the two high-dimensional Last, we also note that the 18-layer plain/residual nets ends.
中文: 这种比较验证了图5(正文)中残余切割的效果被投影所取而代之,人们可以学习极其深厚的系统. 显示时间复杂度和模型大小翻了一番,由于快捷键与"最末"两个高维相接,我们还注意到18层平原/剩余网结束.
<a id="S0160"></a> Source: p.6 S0160
Original: So identity shortcuts lead to more efficient models are comparably accurate (Table 2), but the 18-layer ResNet for the bottleneck designs. converges faster (Fig. 4 right vs. left).
中文: 因此,身份快捷方式导致更有效率的模型比较准确(表2),但用于瓶颈设计的18层ResNet. 汇合速度快(图4右对左).
<a id="S0161"></a> Source: p.6 S0161
Original: When the net is “not 50-layer ResNet: We replace each 2-layer block in the overly deep” (18 layers here), the current SGD solver is still able to find good solutions to the plain net.
中文: 当网络是“不是50层的ResNet:我们替换了过度深处的每个2层的块”(此处有18层)时,目前的SGD解决器仍然能够找到平地网的良好解决办法。
<a id="S0162"></a> Source: p.6 S0162
Original: In this case, the 4Deeper non-bottleneck ResNets (e.g., Fig. 5 left) also gain accuracy ResNet eases the optimization by providing faster converfrom increased depth (as shown on CIFAR-10), but are not as economical gence at the early stage. as the bottleneck ResNets.
中文: 在这种情况下,4Deper非bottleneck ResNets(如图5左出)也获得了精度. ResNet通过提供从增加深度(如CIFAR-10所显示的)更快的汇合器来缓解优化,但在早期阶段却不如经济的汇合. 作为瓶颈ResNets。
<a id="S0163"></a> Source: p.6 S0163
Original: So the usage of bottleneck designs is mainly due to practical considerations.
中文: 因此,瓶颈设计的使用主要出于实际考虑.
<a id="S0164"></a> Source: p.6 S0164
Original: We further note that the degradation problem Identity vs.
中文: 我们还注意到,退化问题与身份问题相对。
<a id="S0165"></a> Source: p.6 S0165
Original: We have shown that of plain nets is also witnessed for the bottleneck designs. 6
中文: 我们已经表明,人们也目睹了普通网的瓶颈设计。 6个
<a id="S0166"></a> Source: p.7 S0166
Original: 34-layer net with this 3-layer bottleneck block, resulting in method error (%) a 50-layer ResNet (Table 1).
中文: 34层网有这个3层的瓶颈块,导致方法错误 (%) 一个50层的ResNet(表1).
<a id="S0167"></a> Source: p.7 S0167
Original: We use option B for increasing Maxout [10] 9.38 dimensions.
中文: 我们使用选项B来增加最大值 [10] 9.38 维度。
<a id="S0168"></a> Source: p.7 S0168
Original: NIN [25] 8.81 101-layer and 152-layer ResNets: We construct 101- DSN [24] 8.22 layer and 152-layer ResNets by using more 3-layer blocks # layers # params (Table 1).
中文: NIN [25] 8.81 101层和152层ResNets:我们通过使用更多的3层块#地层#参数来构建101-DSN [24] 8.22层和152层ResNets(表一).
<a id="S0169"></a> Source: p.7 S0169
Original: Remarkably, although the depth is significantly FitNet [35] 19 2.5M 8.39 increased, the 152-layer ResNet (11.3 billion FLOPs) still Highway [42, 43] 19 2.3M 7.54 (7.72±0.16) has lower complexity than VGG-16/19 nets (15.3/19.6 bil- Highway [42, 43] 32 1.25M 8.80 lion FLOPs).
中文: 值得注意的是,虽然深度显著为FitNet [35] 19 2.5M 8.39增加,但152层的ResNet(113亿FLOPs)仍为高速公路[42,43] 19 2.3M 7.54 (7.72±0.16) 比VGG-16/19网(15.3/19.6 bil-Highway [42,43] 32 1.25M 8.80狮子FLOPs)的复杂度更低.
<a id="S0170"></a> Source: p.7 S0170
Original: ResNet 20 0.27M 8.75 The 50/101/152-layer ResNets are more accurate than ResNet 32 0.46M 7.51 the 34-layer ones by considerable margins (Table 3 and 4).
中文: ResNet 20 0.27M 8.75 50/101/152级的ResNet比ResNet 32 0.46M 7.51级的ResNet更准确,差幅很大(表3和4)。
<a id="S0171"></a> Source: p.7 S0171
Original: ResNet 44 0.66M 7.17 We do not observe the degradation problem and thus en- ResNet 56 0.85M 6.97 joy significant accuracy gains from considerably increased ResNet 110 1.7M 6.43 (6.61±0.16) depth.
中文: ResNet 44 0.66M 7.17 我们没有看到退化问题,因此,ResNet 56 0.85M 6.97的喜悦程度因ResNet 110 1.7M 6.43 (6.61±0.16)深度大幅增加而显著提高。
<a id="S0172"></a> Source: p.7 S0172
Original: The benefits of depth are witnessed for all evaluation ResNet 1202 19.4M 7.93 metrics (Table 3 and 4).
中文: 所有评价都看到深度的好处。
<a id="S0173"></a> Source: p.7 S0173
Original: Classification error on the CIFAR-10 test set.
中文: CIFAR-10测试集的分级错误.
<a id="S0174"></a> Source: p.7 S0174
Original: All meth- Comparisons with State-of-the-art Methods.
中文: 所有冰毒 - 与最新方法的比较。
<a id="S0175"></a> Source: p.7 S0175
Original: In Table 4 ods are with data augmentation.
中文: 在表4中,数据增加。
<a id="S0176"></a> Source: p.7 S0176
Original: For ResNet-110, we run it 5 times we compare with the previous best single-model results. and show “best (mean±std)” as in [43].
中文: 对ResNet-110来说,我们运行它5倍于之前最好的单模型结果. 并显示[43]中“最佳(平均)”字样。
<a id="S0177"></a> Source: p.7 S0177
Original: Our baseline 34-layer ResNets have achieved very competitive accuracy.
中文: 我们的基线34层的ResNet达到了非常有竞争力的准确性。
<a id="S0178"></a> Source: p.7 S0178
Original: Our 152-layer ResNet has a single-model so our residual models have exactly the same depth, width, top-5 validation error of 4.49%.
中文: 我们的152层的ResNet有一个单一的模型,所以我们的剩余模型的深度,宽度完全相同,前5级验证误差为4.49%.
<a id="S0179"></a> Source: p.7 S0179
Original: This single-model result and number of parameters as the plain counterparts. outperforms all previous ensemble results (Table 5).
中文: 这种单模型的结果和参数数作为平面对应. 业绩超过以往所有综艺成果(表5)。
<a id="S0180"></a> Source: p.7 S0180
Original: We We use a weight decay of 0.0001 and momentum of 0.9, combine six models of different depth to form an ensemble and adopt the weight initialization in [13] and BN [16] but (only with two 152-layer ones at the time of submitting). with no dropout.
中文: 我们使用了0.0001的重量衰变和0.9的动力,结合了6个不同深度的模型组成了集合体,并在[13]和BN [16]中采用重量初始化,但(在提交时只有两个152级模型). 没有退学。
<a id="S0181"></a> Source: p.7 S0181
Original: These models are trained with a mini- This leads to 3.57% top-5 error on the test set (Table 5). batch size of 128 on two GPUs.
中文: 这些模型是用迷你 -- 这导致测试集的上57%出错(表5)。 两个GPU上的批量尺寸为128个.
<a id="S0182"></a> Source: p.7 S0182
Original: We start with a learning This entry won the 1st place in ILSVRC 2015. rate of 0.1, divide it by 10 at 32k and 48k iterations, and 4.2.
中文: 我们从学习开始 本条目获得ILSVRC2015年度第1名. 比率为0.1,在32k和48k重迭时除以10;和4.2。
<a id="S0183"></a> Source: p.7 S0183
Original: CIFAR-10 and Analysis terminate training at 64k iterations, which is determined on a 45k/5k train/val split.
中文: CIFAR-10与Analysis终止了64k重复式的训练,这在45k/5k列车/val分出.
<a id="S0184"></a> Source: p.7 S0184
Original: We follow the simple data augmen- We conducted more studies on the CIFAR-10 dataset tation in [24] for training: 4 pixels are padded on each side, [20], which consists of 50k training images and 10k testand a 32 32 crop is randomly sampled from the padded ing images in 10 classes.
中文: 我们遵循简单的数据 ugmen - 我们在[24]中对CIFAR-10数据集tation进行了更多的研究,用于培训:每边有4个像素被加成,[20],由50k训练图像和10k测试组成,从10个班级的加成图像中随机抽取了32个收成.
<a id="S0185"></a> Source: p.7 S0185
Original: We present experiments trained × image or its horizontal flip.
中文: 我们介绍训练成的实验 或横向翻转
<a id="S0186"></a> Source: p.7 S0186
Original: For testing, we only evaluate on the training set and evaluated on the test set.
中文: 为了测试,我们只评价培训单元,评价测试单元。
<a id="S0187"></a> Source: p.7 S0187
Original: Our focus the single view of the original 32 32 image. is on the behaviors of extremely deep networks, but not on × We compare n = 3, 5, 7, 9 , leading to 20, 32, 44, and pushing the state-of-the-art results, so we intentionally use { } 56-layer networks.
中文: 我们关注3232原作的单景. 我们比较n=3,5,7,9, 导致20,32,44, 推动最新结果, 所以我们有意使用{}56层网络。
<a id="S0188"></a> Source: p.7 S0188
Original: Fig. 6 (left) shows the behaviors of the simple architectures as follows. plain nets.
中文: 图6 (左)显示了简单架构的行为如下. 平地网.
<a id="S0189"></a> Source: p.7 S0189
Original: The deep plain nets suffer from increased depth, The plain/residual architectures follow the form in Fig. 3 and exhibit higher training error when going deeper.
中文: 深平地网的深度增加,平地/残存建筑沿图3的形态而来,更深处的训练有误。
<a id="S0190"></a> Source: p.7 S0190
Original: The network inputs are 32 32 images, with phenomenon is similar to that on ImageNet (Fig. 4, left) and × the per-pixel mean subtracted.
中文: 网络输入为32个32个图像,现象类似于ImageNet(图4,左)上的数字,而×每像素平均值被减去.
<a id="S0191"></a> Source: p.7 S0191
Original: The first layer is 3 3 convoon MNIST (see [42]), suggesting that such an optimization × lutions.
中文: 第一层是3 3 Convoon MNIST (见[42]),暗示这种优化×用.
<a id="S0192"></a> Source: p.7 S0192
Original: Then we use a stack of 6n layers with 3 3 convodifficulty is a fundamental problem. × lutions on the feature maps of sizes 32, 16, 8 respectively, Fig. 6 (middle) shows the behaviors of ResNets.
中文: 然后我们用一叠6n的地层 加上3个3个难题 是一个根本的问题。 × 在大小分别为32,16,8的地物图上分别用"图"6(中)来显示ResNets的行为.
<a id="S0193"></a> Source: p.7 S0193
Original: Also { } with 2n layers for each feature map size.
中文: 另外 { } 每个功能地图大小有 2n 层 。
<a id="S0194"></a> Source: p.7 S0194
Original: The numbers of similar to the ImageNet cases (Fig. 4, right), our ResNets filters are 16, 32, 64 respectively.
中文: 与ImageNet案例(图4,右)相类似的数字,我们的ResNets过滤器分别为16,32,64.
<a id="S0195"></a> Source: p.7 S0195
Original: The subsampling is permanage to overcome the optimization difficulty and demon- { } formed by convolutions with a stride of 2.
中文: 子样取法为活活性,以克服优化难度和恶魔-{}由2分步走的进取所形成.
<a id="S0196"></a> Source: p.7 S0196
Original: The network ends strate accuracy gains when the depth increases. with a global average pooling, a 10-way fully-connected We further explore n = 18 that leads to a 110-layer layer, and softmax.
中文: 网络在深度增加时会结束平分精度增益. 全球平均集合,10个全程连接 我们进一步探索 n = 18, 导致一个110层的地层, 和软马克斯。
<a id="S0197"></a> Source: p.7 S0197
Original: There are totally 6n+2 stacked weighted ResNet.
中文: 共有6n+2堆叠式加权ResNet.
<a id="S0198"></a> Source: p.7 S0198
Original: In this case, we find that the initial learning rate layers.
中文: 在这种情况下,我们发现初始学习率层。
<a id="S0199"></a> Source: p.7 S0199
Original: The following table summarizes the architecture: of 0.1 is slightly too large to start converging5.
中文: 下表概括了建筑结构: 0.1个稍大,无法开始汇合5个.
<a id="S0200"></a> Source: p.7 S0200
Original: So we use 0.01 to warm up the training until the training error is below output map size 32×32 16×16 8×8 80% (about 400 iterations), and then go back to 0.1 and con- # layers 1+2n 2n 2n tinue training.
中文: 因此我们使用0.01来给训练取暖,直到训练出错低于输出地图大小32×32 16×16 8×8 80%(约400个迭代),然后回到0.1和con-# 第1+2n 2n 2n 锡克训练.
<a id="S0201"></a> Source: p.7 S0201
Original: The rest of the learning schedule is as done # filters 16 32 64 previously.
中文: 其余的学习时间表和以前一样#过滤16 32 64。
<a id="S0202"></a> Source: p.7 S0202
Original: This 110-layer network converges well (Fig. 6, middle).
中文: 这个110层的网络汇合得很好(图6,中间).
<a id="S0203"></a> Source: p.7 S0203
Original: It has fewer parameters than other deep and thin When shortcut connections are used, they are connected to the pairs of 3 3 layers (totally 3n shortcuts).
中文: 它比起其他深而薄的参数,当使用快捷键连接时,它们被连接到3个3层的对子上(总共是3n快捷键).
<a id="S0204"></a> Source: p.7 S0204
Original: On this 5With an initial learning rate of 0.1, it starts converging (<90% error) × dataset we use identity shortcuts in all cases (i.e., option A), after several epochs, but still reaches similar accuracy. 7
中文: 在此 5 上, 初始学习率为 0.1 , 它开始聚合( < 90% 错误) × 数据集, 我们在所有情况下使用身份快捷键( 即 选项 A) , 在几个纪元之后, 但仍然达到相似的精确度 。 第7条
<a id="S0205"></a> Source: p.8 S0205
Original: 20 10 5 00 1 2 3 4 5 6 iter. (1e4) )%( rorre 20 10 plain-20 5 plain-32 plain-44 plain-56 00 1 2 3 4 5 6 iter. (1e4) )%( rorre 20 ResNet-20 ResNet-32 ResNet-44 ResNet-56 56-layer ResNet-110 20-layer 20-layer 10 110-layer 5 1 0 4 5 6 iter. (1e4) )%( rorre residual-110 residual-1202 Figure 6.
中文: 20 10 00 1 3 4 5 6里特.(1e4)% (ror 20 10平原-20 5平原-32平原-44平原-56平原-2 2 4 5 6里特.(1e4))% (ror 20 ResNet-20 ResNet-32 ResNet-44 ResNet-56 5-56-layer ResNet-110 20-layer 10-layer 5 1 0 4 5-6里特.(1e4))% (ror refore-110剩余-1202 图6.
<a id="S0206"></a> Source: p.8 S0206
Original: Dashed lines denote training error, and bold lines denote testing error.
中文: 虚线表示训练出错,粗线表示测试出错.
<a id="S0207"></a> Source: p.8 S0207
Original: The error of plain-110 is higher than 60% and not displayed.
中文: plain-110的错误高于60%,没有显示.
<a id="S0208"></a> Source: p.8 S0208
Original: Right: ResNets with 110 and 1202 layers. 3 2 1 0 20 40 60 80 100 layer index (sorted by magnitude) dts 3 2 1 0 20 40 60 80 100 layer index (original) plain-20 plain-56 ResNet-20 ResNet-56 ResNet-110 dts plain-20 training data 07+12 07++12 plain-56 ResNet-20 test data VOC 07 test VOC 12 test ResNet-56 ResNet-110 VGG-16 73.2 70.4 ResNet-101 76.4 73.8 Table 7.
中文: 右:有110层和1202层的ResNet. 3, 1 0 20 40 60 80 100层指数(按数量排序) dts. 3, 2 1 0 20 40 60 80 100层指数(原) plain-20 plain-56 ResNet-20 ResNet-56 ResNet-110 dts plain-20培训数据 07+12 07+12 ResNet-20测试数据 VOC 07测试 VOC 12测试 ResNet-56 ResNet-110 VGG-16 73.2 70.4 ResNet-101 76.4 73.8 表7. 按区域分列的
<a id="S0209"></a> Source: p.8 S0209
Original: Object detection mAP (%) on the PASCAL VOC 2007/2012 test sets using baseline Faster R-CNN.
中文: PASCAL VOC 2007/2012测试集上使用基线快取R-CNN的物件检测mAP (%).
<a id="S0210"></a> Source: p.8 S0210
Original: See also Table 10 and 11 for better results. metric mAP@.5 mAP@[.5, .95] VGG-16 41.5 21.2 ResNet-101 48.4 27.2 Figure 7.
中文: 另见表10和11。 r. mAP@.5 mAP@.95] VGG-16 4.1.5 21.2 ResNet-101 48.4 27.2 图7
<a id="S0211"></a> Source: p.8 S0211
Original: Standard deviations (std) of layer responses on CIFAR- 10.
中文: CIFAR-10上的分层响应标准偏差(std).
<a id="S0212"></a> Source: p.8 S0212
Original: The responses are the outputs of each 3×3 layer, after BN and Table 8.
中文: 这些答复是BN和表8之后每3×3层的产出。
<a id="S0213"></a> Source: p.8 S0213
Original: Object detection mAP (%) on the COCO validation set before nonlinearity.
中文: 非线性之前的COCO验证装置上的对象检测 mAP (%).
<a id="S0214"></a> Source: p.8 S0214
Original: Top: the layers are shown in their original using baseline Faster R-CNN.
中文: 顶:用基线快取R-CNN在原地显示地层.
<a id="S0215"></a> Source: p.8 S0215
Original: See also Table 9 for better results. order.
中文: 另见表9。 秩序。 秩序。
<a id="S0216"></a> Source: p.8 S0216
Original: Bottom: the responses are ranked in descending order. have similar training error.
中文: 下行:答复按下行顺序排列. 类似训练错误。
<a id="S0217"></a> Source: p.8 S0217
Original: We argue that this is because of networks such as FitNet [35] and Highway [42] (Table 6), overfitting.
中文: 我们认为,这是由于FitNet[35]和高速公路[42]等网络(表6)的过度适应。
<a id="S0218"></a> Source: p.8 S0218
Original: The 1202-layer network may be unnecessarily yet is among the state-of-the-art results (6.43%, Table 6). large (19.4M) for this small dataset.
中文: 1202-层网络可能不必要地属于最新成果(6.43%,表6)。 大型(19.4M)用于这个小数据集.
<a id="S0219"></a> Source: p.8 S0219
Original: Strong regularization such as maxout [10] or dropout [14] is applied to obtain the Analysis of Layer Responses.
中文: 严格规范化如最大出 [10] 或辍学 [14] 用于获取"地层反应分析".
<a id="S0220"></a> Source: p.8 S0220
Original: Fig. 7 shows the standard best results ([10, 25, 24, 35]) on this dataset.
中文: 图7显示了本数据集的标准最佳结果([10,25,24,35]).
<a id="S0221"></a> Source: p.8 S0221
Original: In this paper, deviations (std) of the layer responses.
中文: 在本文中,层的偏差(std)反应.
<a id="S0222"></a> Source: p.8 S0222
Original: The responses are we use no maxout/dropout and just simply impose regularthe outputs of each 3 3 layer, after BN and before other ization via deep and thin architectures by design, without × nonlinearity (ReLU/addition).
中文: 答案是,我们不使用最大输出/掉出,仅仅将每3个层的输出固定在BN之后,在通过设计深而薄的架构进行其他分解之前,而不使用X非线性(ReLU/添加).
<a id="S0223"></a> Source: p.8 S0223
Original: For ResNets, this analydistracting from the focus on the difficulties of optimizasis reveals the response strength of the residual functions. tion.
中文: 就ResNets而言,从注重Opimizasis的困难这一分析显示剩余功能的反应力。 tion.
<a id="S0224"></a> Source: p.8 S0224
Original: But combining with stronger regularization may im- Fig. 7 shows that ResNets have generally smaller responses prove results, which we will study in the future. than their plain counterparts.
中文: 但是,结合更强有力的规范化,可能会出现图7,显示ResNets的反应一般较小,证明结果,我们将在今后加以研究。 比他们的平凡的对手。
<a id="S0225"></a> Source: p.8 S0225
Original: These results support our basic motivation (Sec.3.1) that the residual functions might 4.3.
中文: 这些结果支持我们的基本动机(第3.1条),即剩余职能可能为4.3。
<a id="S0226"></a> Source: p.8 S0226
Original: Object Detection on PASCAL and MS COCO be generally closer to zero than the non-residual functions.
中文: PASCAL和MS COCO上的对象检测一般比非剩余函数更接近于零.
<a id="S0227"></a> Source: p.8 S0227
Original: Our method has good generalization performance on We also notice that the deeper ResNet has smaller magniother recognition tasks.
中文: 我们的方法有很好的概括性性能 我们还注意到更深的ResNet有较小的放大识别任务.
<a id="S0228"></a> Source: p.8 S0228
Original: Table 7 and 8 show the object detudes of responses, as evidenced by the comparisons among tection baseline results on PASCAL VOC 2007 and 2012 ResNet-20, 56, and 110 in Fig. 7.
中文: 表7和表8显示了答复对象的解析情况,其证据是2007年和2012年的PASCAL VOC和2012年的ResNet-20、56和110图7中的线粒体基线结果的比较。
<a id="S0229"></a> Source: p.8 S0229
Original: We adopt Faster R-CNN [32] as the delayers, an individual layer of ResNets tends to modify the tection method.
中文: 我们采用更快捷的R-CNN [32]作为延时器,一个单层的ResNet倾向于修改齿接法.
<a id="S0230"></a> Source: p.8 S0230
Original: Here we are interested in the improvements signal less. of replacing VGG-16 [41] with ResNet-101.
中文: 在这里,我们对改进信号较少感兴趣。
<a id="S0231"></a> Source: p.8 S0231
Original: The detection Exploring Over 1000 layers.
中文: 探测到1000多层
<a id="S0232"></a> Source: p.8 S0232
Original: We explore an aggressively implementation (see appendix) of using both models is the deep model of over 1000 layers.
中文: 我们探索积极实施(见附录)使用这两个模型,这是1000多层的深层模型。
<a id="S0233"></a> Source: p.8 S0233
Original: We set n = 200 that same, so the gains can only be attributed to better networks. leads to a 1202-layer network, which is trained as described Most remarkably, on the challenging COCO dataset we obabove.
中文: 我们设定的n=200相同,所以收益只能归功于更好的网络. 导致一个1202层的网络, 其培训最显著的描述, 有关挑战性COCO数据集 我们忽略。
<a id="S0234"></a> Source: p.8 S0234
Original: Our method shows no optimization difficulty, and tain a 6.0% increase in COCO’s standard metric (mAP@[.5, this 103-layer network is able to achieve training error .95]), which is a 28% relative improvement.
中文: 我们的方法没有显示优化难度,COCO的标准度量(mAP@[5],这个103层的网络能够实现训练出错.95])增加了6.0%,这是28%的相对改进.
<a id="S0235"></a> Source: p.8 S0235
Original: Its test error is still fairly good solely due to the learned representations. (7.93%, Table 6).
中文: 它的测试错误仍然相当好,完全是由于学到的表述。 (7.93%,表6)。
<a id="S0236"></a> Source: p.8 S0236
Original: Based on deep residual nets, we won the 1st places in But there are still open problems on such aggressively several tracks in ILSVRC & COCO 2015 competitions: Imdeep models.
中文: 在2015年的ILSVRC和COCO比赛中,
<a id="S0237"></a> Source: p.8 S0237
Original: The testing result of this 1202-layer network ageNet detection, ImageNet localization, COCO detection, is worse than that of our 110-layer network, although both and COCO segmentation.
中文: 这种1202-层网络年龄网检测,ImageNet本地化,COCO检测的测试结果比我们110层网络的测试结果要差,尽管两者兼有COCO分化.
<a id="S0238"></a> Source: p.9 S0238
Original: On the number of linear regions of deep neural networks.
中文: 关于深层神经网络的线性区域数量.
<a id="S0239"></a> Source: p.9 S0239
Original: Rectified linear units improve restricted cies with gradient descent is difficult.
中文: 经校正的线性单位改善有梯度回落的受限子是困难的.
<a id="S0240"></a> Source: p.9 S0240
Original: IEEE Transactions on Neural boltzmann machines.
中文: IEEE交易在神经螺栓机上.
<a id="S0241"></a> Source: p.9 S0241
Original: Fisher kernels on visual vocabularies for [2] C. M.
中文: Fisher内核在视觉词汇上为[2] C. M.
<a id="S0242"></a> Source: p.9 S0242
Original: Neural networks for pattern recognition.
中文: 神经网络用于模式识别.
<a id="S0243"></a> Source: p.9 S0243
Original: In CVPR, 2007. university press, 1995. [31] T.
中文: 在CVPR, 2007. 大学出版社, 1995. [31] T.
<a id="S0244"></a> Source: p.9 S0244
Original: Siam, linear transformations in perceptrons.
中文: 暹罗,直觉中的线性转变.
<a id="S0245"></a> Source: p.9 S0245
Original: The devil real-time object detection with region proposal networks.
中文: 魔鬼实时物体检测与区域建议网络.
<a id="S0246"></a> Source: p.9 S0246
Original: In NIPS, is in the details: an evaluation of recent feature encoding methods. 2015.
中文: 在NIPS中,详见:近期特征编码方法的评价. 2015 (英语).
<a id="S0247"></a> Source: p.9 S0247
Original: Zisnetworks on convolutional feature maps. arXiv:1504.06066, 2015. serman.
中文: Zisnetworks on convolutional special mapture. arXiv:1504.0666, 2015 (英语). Serman.
<a id="S0248"></a> Source: p.9 S0248
Original: The Pascal Visual Object Classes (VOC) Challenge.
中文: 帕斯卡尔视觉对象类(VOC)挑战.
<a id="S0249"></a> Source: p.9 S0249
Original: Pattern recognition and neural networks.
中文: 模式识别与神经网络.
<a id="S0250"></a> Source: p.9 S0250
Original: Cambridge pages 303–338, 2010. university press, 1996. [6] S.
中文: 剑桥: 第303–338页. 2010. 大学出版社, 1996. [6]. 师曰.
<a id="S0251"></a> Source: p.9 S0251
Original: Object detection via a multi-region & [35] A.
中文: 通过多区域 & [35] A.
<a id="S0252"></a> Source: p.9 S0252
Original: Gatta, and semantic segmentation-aware cnn model.
中文: Gatta)和语义分解-知觉克能模型.
<a id="S0253"></a> Source: p.9 S0253
Original: Imagenet archies for accurate object detection and semantic segmentation.
中文: 图像网拱门用于精确的物体检测和语义分解.
<a id="S0254"></a> Source: p.9 S0254
Original: In large scale visual recognition challenge. arXiv:1409.0575, 2014.
中文: 在大规模视觉识别挑战中. arXiv:1409.0575, 2014 (英语).
<a id="S0255"></a> Source: p.9 S0255
Original: Understanding the difficulty of training the nonlinear dynamics of learning in deep linear neural networks. deep feedforward neural networks.
中文: 了解深线性神经网络中学习的非线性动力学培训难度. 深源神经网络.
<a id="S0256"></a> Source: p.9 S0256
Original: In AISTATS, 2010. arXiv:1312.6120, 2013. [10] I. J.
中文: 在AISTATS, 2010. arXiv:1312.6120, 2013 [10] I. J.
<a id="S0257"></a> Source: p.9 S0257
Original: Accelerated gradient descent by factor-centering Y.
中文: 由因子中心Y加速梯度下降.
<a id="S0258"></a> Source: p.9 S0258
Original: Maxout networks. arXiv:1302.4389, 2013. decomposition.
中文: Maxout网络. arXiv:1302.4389 2013. 分解.
<a id="S0259"></a> Source: p.9 S0259
Original: Convolutional neural networks at constrained time [39] N. N.
中文: 受限时间的革命神经网络[39] (原始内容存档于2013-10-12). N. N.
<a id="S0260"></a> Source: p.9 S0260
Original: Centering neural network gradient factors.
中文: 中心神经网络梯度因子.
<a id="S0261"></a> Source: p.9 S0261
Original: Neural Networks: Tricks of the Trade, pages 207–226.
中文: 神经网络:贸易的诡计, 第207–226页.
<a id="S0262"></a> Source: p.9 S0262
Original: Spatial pyramid pooling in deep 1998. convolutional networks for visual recognition.
中文: 1998年深处的空间金字塔集合. 视觉识别的革命网络.
<a id="S0263"></a> Source: p.9 S0263
Original: Overfeat: Integrated recognition, localization and detection Surpassing human-level performance on imagenet classification.
中文: 过度feat:综合识别,本地化和检测 Surpassing 人类层面在图像网分类上的性能.
<a id="S0264"></a> Source: p.9 S0264
Original: Very deep convolutional networks [14] G. E.
中文: 非常深刻的革命网络 [14] G. E.
<a id="S0265"></a> Source: p.9 S0265
Original: Sutskever, and for large-scale image recognition.
中文: Sutskever, 以及大规模图像识别。
<a id="S0266"></a> Source: p.9 S0266
Original: Improving neural networks by preventing co- [42] R. K.
中文: 通过共同防止[42] R. K.
<a id="S0267"></a> Source: p.9 S0267
Original: Highway networks. adaptation of feature detectors. arXiv:1207.0580, 2012. arXiv:1505.00387, 2015. [15] S.
中文: 高速公路网. 改装地物探测器。 arXiv:1207.0580, 2012. arXiv:1505.00387, 2015 [15] (中文(中国大陆) ). 师曰.
<a id="S0268"></a> Source: p.9 S0268
Original: Training very deep computation, 9(8):1735–1780, 1997. networks. 1507.06228, 2015. [16] S.
中文: 训练非常深入的计算,9(8):1735–1780,1997. Network. 1507.06228, 2015 [16]. 师曰.
<a id="S0269"></a> Source: p.9 S0269
Original: Batch normalization: Accelerating deep [44] C.
中文: 批量正常化:加速深度[44] C.
<a id="S0270"></a> Source: p.9 S0270
Original: Ernetwork training by reducing internal covariate shift.
中文: 通过减少内部同位变换来进行网络培训.
<a id="S0271"></a> Source: p.9 S0271
Original: Fast surface interpolation using hierarchical basis func- [18] H.
中文: 使用分级基础真菌-[18] H.
<a id="S0272"></a> Source: p.9 S0272
Original: Aggregating local image descriptors into compact codes. [46] R.
中文: 将本地图像描述符汇总为紧凑代码。 [46] R. (英语).
<a id="S0273"></a> Source: p.9 S0273
Original: Locally adapted hierarchical basis preconditioning.
中文: 地方调整分级基础的先决条件。
<a id="S0274"></a> Source: p.9 S0274
Original: Caffe: Convolutional architecture for tic gradient towards second-order methods–backpropagation learnfast feature embedding. arXiv:1408.5093, 2014. ing with transformations in nonlinearities.
中文: Caffe: 向二等方法的梯度进化结构 – 背向传播学习快感特征嵌入 。 arXiv:1408.5093, 2014. 与非线性变换相通.
<a id="S0275"></a> Source: p.9 S0275
Original: Learning multiple layers of features from tiny im- Processing, 2013. ages.
中文: 从微小的不加工中学习多层特征, 2013. ages.
<a id="S0276"></a> Source: p.9 S0276
Original: VLFeat: An open and portable library [21] A.
中文: VLFeat:一个开放而可移植的图书馆 [21] A.
<a id="S0277"></a> Source: p.9 S0277
Original: Imagenet classification of computer vision algorithms, 2008. with deep convolutional neural networks.
中文: 计算机视觉算法的图像网分类, 2008. 有深演化神经网络.
<a id="S0278"></a> Source: p.9 S0278
Original: Modern applied statistics with s-plus. [22] Y.
中文: 具有 " s-plus " 的现代应用统计[22] 没错
<a id="S0279"></a> Source: p.9 S0279
Original: Backpropagation applied to hand- [50] M. D.
中文: 反扩散适用于手-[50] M. D.
<a id="S0280"></a> Source: p.9 S0280
Original: Visualizing and understanding convoluwritten zip code recognition.
中文: 可视化和理解可通缩码识别.
<a id="S0281"></a> Source: p.9 S0281
Original: Neural computation, 1989. tional neural networks.
中文: 神经计算 1989. 通心神经网络.
<a id="S0282"></a> Source: p.9 S0282
Original: In Neural Networks: Tricks of the Trade, pages 9–50.
中文: 《神经网络:贸易的花招》,第9-50页。
<a id="S0283"></a> Source: p.9 S0283
Original: Deeplysupervised nets. arXiv:1409.5185, 2014. [25] M.
中文: 深度监管网. arXiv:1409.5185, 2014. [25] M.
<a id="S0284"></a> Source: p.9 S0284
Original: Network in network. arXiv:1312.4400, 2013. [26] T.-Y.
中文: 网易. arXiv:1312.4400, 2013. [26] T.-Y.
<a id="S0285"></a> Source: p.9 S0285
Original: Microsoft COCO: Common objects in context.
中文: 微软COCO:上下文中的常见对象.
<a id="S0286"></a> Source: p.9 S0286
Original: Fully convolutional networks for semantic segmentation.
中文: 充分演化语义分化网络.
<a id="S0287"></a> Source: p.10 S0287
Original: Object Detection Baselines 8 images (i.e., 1 per GPU) and the Fast R-CNN step has a mini-batch size of 16 images.
中文: 对象探测基线为8个图像(即每GPU1个),快取R-CNN步骤有16个图像的小型批量大小.
<a id="S0288"></a> Source: p.10 S0288
Original: The RPN step and Fast R- In this section we introduce our detection method based CNN step are both trained for 240k iterations with a learnon the baseline Faster R-CNN [32] system.
中文: RPN步骤和快活R-在本节中,我们采用基于CNN步骤的探测方法,对240克重迭进行了训练,并学习了基线快活R-CNN[32]系统。
<a id="S0289"></a> Source: p.10 S0289
Original: The models are ing rate of 0.001 and then for 80k iterations with 0.0001. initialized by the ImageNet classification models, and then Table 8 shows the results on the MS COCO validation fine-tuned on the object detection data.
中文: 型号为0.001转速,再以0.0001转速80克. 由ImageNet分类模型初始化,再由表8显示在对象检测数据上微调的MS COCO验证结果.
<a id="S0290"></a> Source: p.10 S0290
Original: ResNet-101 has a 6% increase of mAP@[.5, .95] over mented with ResNet-50/101 at the time of the ILSVRC & VGG-16, which is a 28% relative improvement, solely con- COCO 2015 detection competitions. tributed by the features learned by the better network.
中文: ResNet-101比在ILSVRC & VGG-16出道时使用ResNet-50/101进行修饰时的mAP@[5.95]增加了6%,相对改进了28%,仅是Con-CO 2015检测比赛. 以更好的网络所学到的特色为荣.
<a id="S0291"></a> Source: p.10 S0291
Original: Re- Unlike VGG-16 used in [32], our ResNet has no hidden markably, the mAP@[.5, .95]’s absolute increase (6.0%) is fc layers.
中文: 与[32]中使用的VGG-16不同的是,我们的ResNet没有隐藏的可标可标,mAP@[5.95]的绝对增加(6.0%)为fc分层.
<a id="S0292"></a> Source: p.10 S0292
Original: We adopt the idea of “Networks on Conv feanearly as big as mAP@.5’s (6.9%).
中文: 我们采纳了“与mAP(6.9%)一样远大节日网络”的想法。
<a id="S0293"></a> Source: p.10 S0293
Original: This suggests that a ture maps” (NoC) [33] to address this issue.
中文: 这表明,需要绘制出一张图来解决这一问题”(NoC)[33]。
<a id="S0294"></a> Source: p.10 S0294
Original: We compute deeper network can improve both recognition and localizathe full-image shared conv feature maps using those laytion. ers whose strides on the image are no greater than 16 pixels (i.e., conv1, conv2 x, conv3 x, and conv4 x, totally 91 conv B.
中文: 我们计算更深层的网络 既能提高识别度 也能利用这些图谱 使全图像共享的曲线图更加本地化 其图像上出行速度不大于16像素的ers(即:conv1,conv2 x,conv3 x;和conv4 x,完全91 conv B.
<a id="S0295"></a> Source: p.10 S0295
Original: Object Detection Improvements layers in ResNet-101; Table 1).
中文: ResNet-101中的对象检测改进层;表1。
<a id="S0296"></a> Source: p.10 S0296
Original: We consider these layers as analogous to the 13 conv layers in VGG-16, and by doing For completeness, we report the improvements made for so, both ResNet and VGG-16 have conv feature maps of the the competitions.
中文: 我们认为这些地层与VGG-16中的13个凸起地层相似,为了完整起见,我们报告为此所做的改进,ResNet和VGG-16都有凸起地貌的竞赛图.
<a id="S0297"></a> Source: p.10 S0297
Original: These improvements are based on deep same total stride (16 pixels).
中文: 这些改进基于深度相同的全步(16像素).
<a id="S0298"></a> Source: p.10 S0298
Original: These layers are shared by a features and thus should benefit from residual learning. region proposal network (RPN, generating 300 proposals) [32] and a Fast R-CNN detection network [7].
中文: 这些层层有共同的特征,因此应受益于剩余学习。 区域提案网络(RPN,生成300个提案)[32]和一个快速R-CNN检测网络[7]。
<a id="S0299"></a> Source: p.10 S0299
Original: RoI pool- MS COCO ing [7] is performed before conv5 1.
中文: RoI 池 -- MS COCO ing [7] 被执行于 conv5 1.
<a id="S0300"></a> Source: p.10 S0300
Original: Our box refinement partially follows the itfeature, all layers of conv5 x and up are adopted for each erative localization in [6].
中文: 我们的盒子精炼部分地遵循了它的特异性,所有层的凸起5 x 和上层被采用在[6]中每个被去除的地方化.
<a id="S0301"></a> Source: p.10 S0301
Original: In Faster R-CNN, the final output region, playing the roles of VGG-16’s fc layers.
中文: 在更快的R-CNN中,最终输出区域,扮演了VGG-16的fc层角色.
<a id="S0302"></a> Source: p.10 S0302
Original: The final is a regressed box that is different from its proposal box.
中文: 最后是与其建议框不同的一个倒退的方框。
<a id="S0303"></a> Source: p.10 S0303
Original: So classification layer is replaced by two sibling layers (classi- for inference, we pool a new feature from the regressed box fication and box regression [7]). and obtain a new classification score and a new regressed For the usage of BN layers, after pre-training, we com- box.
中文: 因此分级层被两个分级层所取而代之(classi-用于推论,我们从后退的相框虚构和相框回归[7]中汇集出一个新的特征). 并获得一个新的分类分数 和一个新的退步 对于BN层的使用, 经过前期训练,我们 com-box。
<a id="S0304"></a> Source: p.10 S0304
Original: We combine these 300 new predictions with the origpute the BN statistics (means and variances) for each layer inal 300 predictions.
中文: 我们把这300个新的预测与每个层的300个预测的BN统计(表示和差异)合并。
<a id="S0305"></a> Source: p.10 S0305
Original: Non-maximum suppression (NMS) is on the ImageNet training set.
中文: 非最大压制(NMS)在ImageNet培训集上.
<a id="S0306"></a> Source: p.10 S0306
Original: Then the BN layers are fixed applied on the union set of predicted boxes using an IoU during fine-tuning for object detection.
中文: 然后,BN层被固定在预想的组合箱上,在进行物体检测的微调时使用IoU.
<a id="S0307"></a> Source: p.10 S0307
Original: As such, the BN threshold of 0.3 [8], followed by box voting [6].
中文: 因此,BN的门槛为0.3 [8],然后是票箱投票[6].
<a id="S0308"></a> Source: p.10 S0308
Original: Box relayers become linear activations with constant offsets and finement improves mAP by about 2 points (Table 9). scales, and BN statistics are not updated by fine-tuning.
中文: 箱式中继器成为有常数相抵的线性活化器,而精细使mAP改进了约2分(表9). 缩放,不通过微调更新BN统计数据。
<a id="S0309"></a> Source: p.10 S0309
Original: We combine global context in the Fast fix the BN layers mainly for reducing memory consumption R-CNN step.
中文: 我们在"快报"中结合了全球背景来修复BN层,主要是为了减少内存消耗R-CNN步骤.
<a id="S0310"></a> Source: p.10 S0310
Original: Given the full-image conv feature map, we in Faster R-CNN training. pool a feature by global Spatial Pyramid Pooling [12] (with PASCAL VOC a “single-level” pyramid) which can be implemented as Following [7, 32], for the PASCAL VOC 2007 test set, “RoI” pooling using the entire image’s bounding box as the we use the 5k trainval images in VOC 2007 and 16k train- RoI.
中文: 鉴于全相机凸起的地貌图,我们接受更快的R-CNN训练. 集合全球空间金字塔集合12的功能,可被执行为:为PASCAL VOC 2007测试集,"RoI"集合使用整个图像的边框,作为我们在VOC 2007和16k列车中使用5k列车的图像. 罗
<a id="S0311"></a> Source: p.10 S0311
Original: This pooled feature is fed into the post-RoI layers to val images in VOC 2012 for training (“07+12”).
中文: 这一集合功能被输入到国际公路运输之后的地层到VOC 2012年的val图像中用于培训(“07+12”).
<a id="S0312"></a> Source: p.10 S0312
Original: For the obtain a global context feature.
中文: 为了获得全球背景特征。
<a id="S0313"></a> Source: p.10 S0313
Original: This global feature is con- PASCAL VOC 2012 test set, we use the 10k trainval+test catenated with the original per-region feature, followed by images in VOC 2007 and 16k trainval images in VOC 2012 the sibling classification and box regression layers.
中文: 这个全球特征是Con-PASCAL VOC 2012测试集,我们使用10k列车val+测试被以最初的每个区域特征进行编目,然后是VOC 2007的图像和VOC 2012的16k列车val图像,即兄弟姐妹分类和箱回归层.
<a id="S0314"></a> Source: p.10 S0314
Original: The hyper-parameters for train- new structure is trained end-to-end.
中文: 列车的超参数 - 新结构是训练有素的端到端.
<a id="S0315"></a> Source: p.10 S0315
Original: Global context iming Faster R-CNN are the same as in [32].
中文: " 快速R-CNN " 的全球背景与[32]相同。
<a id="S0316"></a> Source: p.10 S0316
Original: Table 7 shows proves mAP@.5 by about 1 point (Table 9). the results.
中文: 表7显示,mAP@5用大约1分(表9)。 结果。
<a id="S0317"></a> Source: p.10 S0317
Original: ResNet-101 improves the mAP by >3% over Multi-scale testing.
中文: ResNet-101比多尺度测试改进了>3%的mAP.
<a id="S0318"></a> Source: p.10 S0318
Original: In the above, all results are obtained by VGG-16.
中文: 在上述,所有结果由VGG-16获得.
<a id="S0319"></a> Source: p.10 S0319
Original: This gain is solely because of the improved feasingle-scale training/testing as in [32], where the image’s tures learned by ResNet. shorter side is s = 600 pixels.
中文: 这一增益完全是因为改进了Feasingle规模的训练/测试,例如[32],ResNet在其中学到了图像的花纹. 更短的一面是 s = 600 像素.
<a id="S0320"></a> Source: p.10 S0320
Original: Multi-scale training/testing MS COCO has been developed in [12, 7] by selecting a scale from a The MS COCO dataset [26] involves 80 object cate- feature pyramid, and in [33] by using maxout layers.
中文: 通过从一个MS COCO数据集中选择一个比例尺[26],在[12,7]中开发了多种规模的训练/测试MS COCO,在[33]中开发了80个对象的克托克特征金字塔,在[33]中使用了最大出地层.
<a id="S0321"></a> Source: p.10 S0321
Original: We evaluate the PASCAL VOC metric (mAP @ our current implementation, we have performed multi-scale IoU = 0.5) and the standard COCO metric (mAP @ IoU = testing following [33]; we have not performed multi-scale .5:.05:.95).
中文: 我们评价PASCAL VOC 度量衡(mAP@我们目前的执行,我们进行了多级IoU=0.5)和标准COCO度量衡(mAP@IoU=在[33]之后的测试);我们没有进行多级:5.05:95。
<a id="S0322"></a> Source: p.10 S0322
Original: We use the 80k images on the train set for train- training because of limited time.
中文: 我们使用火车上80克的图像 用于火车训练 因为时间有限
<a id="S0323"></a> Source: p.10 S0323
Original: In addition, we have pering and the 40k images on the val set for evaluation.
中文: 此外,我们还有 浏览和40克的图像 在val的一套评估。
<a id="S0324"></a> Source: p.10 S0324
Original: Our formed multi-scale testing only for the Fast R-CNN step detection system for COCO is similar to that for PASCAL (but not yet for the RPN step).
中文: 我们仅对COCO的快取R-CNN阶梯检测系统形成多尺度测试,与PASCAL类似(但尚未对RPN阶梯).
<a id="S0325"></a> Source: p.10 S0325
Original: We train the COCO models with an 8-GPU imple- compute conv feature maps on an image pyramid, where the mentation, and thus the RPN step has a mini-batch size of image’s shorter sides are s 200, 400, 600, 800, 1000 . ∈ { } 10
中文: 我们在图像金字塔上用8-GPU imple-compute Conv 特征图来训练COCO模型,其中的修饰, RPN 步骤有一个小批量的图像短边是200、400、600、800、1000。
<a id="S0326"></a> Source: p.11 S0326
Original: training data COCO train COCO trainval test data COCO val COCO test-dev mAP @.5 @[.5, .95] @.5 @[.5, .95] baseline Faster R-CNN (VGG-16) 41.5 21.2 baseline Faster R-CNN (ResNet-101) 48.4 27.2 +box refinement 49.9 29.9 +context 51.1 30.0 53.3 32.2 +multi-scale testing 53.8 32.5 55.7 34.9 ensemble 59.0 37.4 Table 9.
中文: COCO列车COCO列车列车试验数据 COCO val CO 测试-dev mAP@ 5. @ 5..95] @ 5. 基线 更快 R-CNN (VGG-16) 41.5 21.2 基线 更快 R-CNN (ResNet-101) 48.2 +箱型改进 49.9 29.9 +箱型改进 51.1 30.0 53.3 32.2 +多尺度测试 53.8 32.5 55.7 34.9 综艺 59.0 37.4 表9. 按部门开列的所需资源
<a id="S0327"></a> Source: p.11 S0327
Original: Object detection improvements on MS COCO using Faster R-CNN and ResNet-101. system net data mAP areo bike bird boat bottle bus car cat chair cow table dog horse mbike person plant sheep sofa train tv baseline VGG-16 07+12 73.2 76.5 79.0 70.9 65.5 52.1 83.1 84.7 86.4 52.0 81.9 65.7 84.8 84.6 77.5 76.7 38.8 73.6 73.9 83.0 72.6 baseline ResNet-101 07+12 76.4 79.8 80.7 76.2 68.3 55.9 85.1 85.3 89.8 56.7 87.8 69.4 88.3 88.9 80.9 78.4 41.7 78.6 79.8 85.3 72.0 baseline+++ ResNet-101 COCO+07+12 85.6 90.0 89.6 87.8 80.8 76.1 89.9 89.9 89.6 75.5 90.0 80.7 89.6 90.3 89.1 88.7 65.4 88.1 85.6 89.0 86.8 Table 10.
中文: MS COCO使用更快捷的R-CNN和ResNet-101. 系统网数据mAP为单车鸟船瓶装活活禽车车厢车厢车厢车厢牛座马匹mbike人厂羊沙发列车tv基线VGG-16 07+12 73.2 76.579.0 70.9 65.5 52.1 83.1 84.7 86.4 52.0 65.7 84.8 77.6 73.9 83.0 72.6 基线ResNet-101 07+12 76.4 79.8 80.7 76.2 68.3 55.9 85.1 89.8 87.8 87.8 69.4 88.3 88.9 80.9 78.4 7 78.6 79.8 85.3 72.0 基线*** ResNet-101 COCO+07+12 85.6 90. 89.6 87.8 76.1 89.9 89.9 89.6 89.5 90.80.7 89.3 89.7 65.4 85.1 85.6 89.8 表10 10.
<a id="S0328"></a> Source: p.11 S0328
Original: Detection results on the PASCAL VOC 2007 test set.
中文: PASCAL VOC 2007测试集的检测结果.
<a id="S0329"></a> Source: p.11 S0329
Original: The baseline is the Faster R-CNN system.
中文: 基线是更快的R-CNN系统.
<a id="S0330"></a> Source: p.11 S0330
Original: The system “baseline+++” include box refinement, context, and multi-scale testing in Table 9. system net data mAP areo bike bird boat bottle bus car cat chair cow table dog horse mbike person plant sheep sofa train tv baseline VGG-16 07++12 70.4 84.9 79.8 74.3 53.9 49.8 77.5 75.9 88.5 45.6 77.1 55.3 86.9 81.7 80.9 79.6 40.1 72.6 60.9 81.2 61.5 baseline ResNet-101 07++12 73.8 86.5 81.6 77.2 58.0 51.0 78.6 76.6 93.2 48.6 80.4 59.0 92.1 85.3 84.8 80.7 48.1 77.3 66.5 84.7 65.6 baseline+++ ResNet-101 COCO+07++12 83.8 92.1 88.4 84.8 75.9 71.4 86.3 87.8 94.2 66.8 89.4 69.2 93.9 91.9 90.9 89.6 67.9 88.2 76.8 90.3 80.0 Table 11.
中文: “基线”系统包括表9中的箱型改进、上下文和多尺度测试。 系统网数据为:mAP eo 自行车鸟船 瓶装公交车 奶牛座车 狗马 摩托车人 花羊沙发火车 tv 基线 VGG-16 07++12 70.4 84.979.8 74.3 53.9 49.8 77.5 75.9 88.5 80.6 77.6 80.6 60.9 81.6.9 81.5 基线 ResNet-101 07++12 73.8 86.5 81.6 77.2 58.0 51.0 51.0 7 76.6 93.6 48.6 80.4 59.0 92.1 85.3 84.8 80.7 基线 ResNet-101 CO+07++12 83.8 92.1 88.4 75.9 71.4 86.3 87.8 94.2 66.8 89.4 93.9 90.9 90.9 89.6 67.9 88.2 80.3 表11。
<a id="S0331"></a> Source: p.11 S0331
Original: Detection results on the PASCAL VOC 2012 test set (http://host.robots.ox.ac.uk:8080/leaderboard/ displaylb.php?challengeid=11&compid=4).
中文: PASCAL VOC 2012测试集的检测结果(http://host.robots.ox.ac.uk:8080/leaderboard/extlb.php? challengeid=11&compid=4).
<a id="S0332"></a> Source: p.11 S0332
Original: The baseline is the Faster R-CNN system.
中文: 基线是更快的R-CNN系统.
<a id="S0333"></a> Source: p.11 S0333
Original: The system “baseline+++” include box refinement, context, and multi-scale testing in Table 9.
中文: 系统“基线”包括表9中的方框改进、上下文和多尺度测试。
<a id="S0334"></a> Source: p.11 S0334
Original: We select two adjacent scales from the pyramid following val2 test [33].
中文: 我们从金字塔上选取两个相邻的天平 经过val2测试[33].
<a id="S0335"></a> Source: p.11 S0335
Original: RoI pooling and subsequent layers are performed on GoogLeNet [44] (ILSVRC’14) - 43.9 the feature maps of these two scales [33], which are merged our single model (ILSVRC’15) 60.5 58.8 by maxout as in [33].
中文: RoI集和随后的地层在GoogleNet44上进行 -- -- 43.9这两个比例尺的地物图[33],它们被我们的单一模型(ILSVRC ' 15)合并为60.5 58.8, 和[33]中的最大输出。
<a id="S0336"></a> Source: p.11 S0336
Original: Multi-scale testing improves the mAP our ensemble (ILSVRC’15) 63.6 62.1 by over 2 points (Table 9).
中文: 多尺度测试将我们的综艺节目(ILSVRC ' 15)改进了63.6 62.1分以上(表9)。
<a id="S0337"></a> Source: p.11 S0337
Original: Our results (mAP, %) on the ImageNet detection dataset.
中文: 我们在ImageNet检测数据集上的结果(mAP,%).
<a id="S0338"></a> Source: p.11 S0338
Original: Next we use the 80k+40k trainval set Our detection system is Faster R-CNN [32] with the improvements for training and the 20k test-dev set for evaluation.
中文: 接下来我们使用80k+40k列车的列车设置 我们的探测系统是更快捷的R-CNN [32],其改进用于训练,20k测试-dev用于评价.
<a id="S0339"></a> Source: p.11 S0339
Original: The test- in Table 9, using ResNet-101. dev set has no publicly available ground truth and the result is reported by the evaluation server.
中文: 表9中的测试使用ResNet-101. Dev set没有公开的地面真相,结果由评价服务器报告.
<a id="S0340"></a> Source: p.11 S0340
Original: Under this setting, the we achieve 85.6% mAP on PASCAL VOC 2007 (Table 10) results are an mAP@.5 of 55.7% and an mAP@[.5, .95] of and 83.8% on PASCAL VOC 2012 (Table 11)6.
中文: 在这一背景下,我们在2007年PASCAL VOC上实现85.6%的mAP(表10)结果为55.7%的mAP和2012年PASCAL VOC上达到83.8%的mAP(表116)。
<a id="S0341"></a> Source: p.11 S0341
Original: This is our single-model result. on PASCAL VOC 2012 is 10 points higher than the previ- Ensemble.
中文: 这是我们单一模式的结果。 在PASCAL VOC 2012上,比前作"前作"高出10分.
<a id="S0342"></a> Source: p.11 S0342
Original: In Faster R-CNN, the system is designed to learn ous state-of-the-art result [6]. region proposals and also object classifiers, so an ensemble ImageNet Detection can be used to boost both tasks.
中文: 在更快速的R-CNN中,系统的设计旨在学习ous最先进的结果[6]. 区域提案和对象分类器,因此可以使用一个聚合图像网络检测器来推进这两个任务.
<a id="S0343"></a> Source: p.11 S0343
Original: We use an ensemble for The ImageNet Detection (DET) task involves 200 object proposing regions, and the union set of proposals are procategories.
中文: 我们使用一个“图像网络检测”任务组合,涉及200个对象提议区域,而联盟的一组提议是亲类。
<a id="S0344"></a> Source: p.11 S0344
Original: Our cessed by an ensemble of per-region classifiers.
中文: 我们被每个地区分类的 组合失败了
<a id="S0345"></a> Source: p.11 S0345
Original: Table 9 object detection algorithm for ImageNet DET is the same shows our result based on an ensemble of 3 networks.
中文: 表9 ImageNet DET的物体检测算法相同,显示了我们基于由3个网络组成的组合的结果.
<a id="S0346"></a> Source: p.11 S0346
Original: The networks are premAP is 59.0% and 37.4% on the test-dev set.
中文: 网络预告为59.0%,测试-dev集为37.4%.
<a id="S0347"></a> Source: p.11 S0347
Original: This result trained on the 1000-class ImageNet classification set, and won the 1st place in the detection task in COCO 2015. are fine-tuned on the DET data.
中文: 这个结果在1000级的ImageNet分类集上接受了培训,并在COCO 2015的检测任务上获得了第1名. 对DET数据进行了微调。
<a id="S0348"></a> Source: p.11 S0348
Original: We split the validation set PASCAL VOC into two parts (val1/val2) following [8].
中文: 我们在[8]后将验证装置PASCAL VOC分为两部分(val1/val2)。
<a id="S0349"></a> Source: p.11 S0349
Original: We fine-tune the We revisit the PASCAL VOC dataset based on the above detection models using the DET training set and the val1 model.
中文: 我们利用DET培训集和val1模型,根据上述检测模型,对PASCAL VOC数据集进行微调。
<a id="S0350"></a> Source: p.11 S0350
Original: With the single model on the COCO dataset (55.7% set.
中文: 以COCO数据集的单模型为主(55.7%集.
<a id="S0351"></a> Source: p.11 S0351
Original: We do not use other mAP@.5 in Table 9), we fine-tune this model on the PAS- ILSVRC 2015 data.
中文: 在表9中,我们不使用其他mAP@5,我们在PAS-ILSVRC2015数据上微调了这一模式。
<a id="S0352"></a> Source: p.11 S0352
Original: Our single model with ResNet-101 has CAL VOC sets.
中文: 我们与ResNet-101的单一型号有CAL VOC集.
<a id="S0353"></a> Source: p.11 S0353
Original: The improvements of box refinement, con- 6http://host.robots.ox.ac.uk:8080/anonymous/3OJ4OJ.html, text, and multi-scale testing are also adopted.
中文: 改进的盒子,Con... 6http://host.robots.ox.ac.uk:8080/anonymous/3OJ4OJ.html,文本,并采用多尺度测试.
<a id="S0354"></a> Source: p.12 S0354
Original: LOC LOC LOC error classification top-5 LOC error top-5 localization err testing method method network on GT CLS network on predicted CLS val test VGG’s [41] VGG-16 1-crop 33.1 [41] OverFeat [40] (ILSVRC’13) 30.0 29.9 RPN ResNet-101 1-crop 13.3 RPN ResNet-101 dense 11.7 GoogLeNet [44] (ILSVRC’14) - 26.7 RPN ResNet-101 dense ResNet-101 14.4 VGG [41] (ILSVRC’14) 26.9 25.3 RPN+RCNN ResNet-101 dense ResNet-101 10.6 ours (ILSVRC’15) 8.9 9.0 RPN+RCNN ensemble dense ensemble 8.9 Table 14.
中文: LOC LOC 错误分类 上-5 LOC 错误分类 上-5 LOC 错误本地化 GT CLS 网络上关于预测 CLS val 测试 VGG的测试方法网 [41] VGG-16 1-crop 33.1 [41] OverFeat [40] (ILSVRC ' 13] 30.0 29.9 RPN ResNet-101 1-crop 13.3 RPN ResNet-101 稠密 11.7 GoogleNet [44] (ILSVRC ' 14) - 26.7 RPN ResNet-101 稠密ResNet-101 稠密ResNet-101 我们的测试方法网 [41] (ILSVRC ' 14) 26.9 25.3 RPN+RC NNN 稠密ResNet-101 10.6 (ILSVRC ' 15) 8.9.0 RPN 稠密共 8.9 表 14.
<a id="S0355"></a> Source: p.12 S0355
Original: Comparisons of localization error (%) on the ImageNet Table 13.
中文: ImageNet表13对本地化错误(%)的比较.
<a id="S0356"></a> Source: p.12 S0356
Original: Localization error (%) on the ImageNet validation.
中文: 在ImageNet验证上本地化出错 (%).
<a id="S0357"></a> Source: p.12 S0357
Original: In dataset with state-of-the-art methods. the column of “LOC error on GT class” ([41]), the ground truth class is used.
中文: 在具有最先进方法的数据集中. “GT类的LOC出错”栏([41]),使用了地面真理类。
<a id="S0358"></a> Source: p.12 S0358
Original: In the “testing” column, “1-crop” denotes testing on a center crop of 224×224 pixels, “dense” denotes dense (fully ports a center-crop error of 33.1% (Table 13) using ground convolutional) and multi-scale testing. truth classes.
中文: 在“试验”一栏中,“1-作物”表示对224×224像素的中心作物进行试验,“密集”表示密集(完全用地演化法将中心作物误差为33.1%(表13))和多尺度试验。 真理课。
<a id="S0359"></a> Source: p.12 S0359
Original: Under the same setting, our RPN method using ResNet-101 net significantly reduces the center-crop error to 13.3%.
中文: 在同一设定下,我们使用ResNet-101净值的RPN方法将中心作物误差大幅降低到13.3%.
<a id="S0360"></a> Source: p.12 S0360
Original: This comparison demonstrates the excellent 58.8% mAP and our ensemble of 3 models has 62.1% mAP performance of our framework.
中文: 这一比较表明,我们框架的58.8%的MAP和我们3个模型的组合有62.1%的MAP性能。
<a id="S0361"></a> Source: p.12 S0361
Original: With dense (fully convoluon the DET test set (Table 12).
中文: 密度(完全压缩DET测试集(表12)。
<a id="S0362"></a> Source: p.12 S0362
Original: This result won the 1st place tional) and multi-scale testing, our ResNet-101 has an error in the ImageNet detection task in ILSVRC 2015, surpassing of 11.7% using ground truth classes.
中文: 这个结果赢得了第1位的tional)和多尺度测试,我们的ResNet-101在ILSVRC2015年的ImageNet检测任务中有一个出错,使用地面真人类超过了11.7%.
<a id="S0363"></a> Source: p.12 S0363
Original: Using ResNet-101 for the second place by 8.5 points (absolute). predicting classes (4.6% top-5 classification error, Table 4), the top-5 localization error is 14.4%. C.
中文: 使用ResNet-101获得8.5分的第二名(绝对分). 预测类别(4.6%为上至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下至下 C C.
<a id="S0364"></a> Source: p.12 S0364
Original: ImageNet Localization The above results are only based on the proposal network (RPN) in Faster R-CNN [32].
中文: 图像网络本地化 以上结果仅以"快快R-CNN"[32]中的建议网络(RPN)为基础.
<a id="S0365"></a> Source: p.12 S0365
Original: One may use the detection The ImageNet Localization (LOC) task [36] requires to network (Fast R-CNN [7]) in Faster R-CNN to improve the classify and localize the objects.
中文: 人们可以使用检测图像网本地化(LOC)任务[36]要求快速R-CNN[7]在快速R-CNN中联网来改进对象的分类和本地化.
<a id="S0366"></a> Source: p.12 S0366
Original: But we notice that on this dataset, one image usually assume that the image-level classifiers are first adopted for contains a single dominate object, and the proposal regions predicting the class labels of an image, and the localizahighly overlap with each other and thus have very similar tion algorithm only accounts for predicting bounding boxes RoI-pooled features.
中文: 但是我们注意到,在这个数据集中,一个图像通常假设图像级别分类器首先被采用为包含一个单一的主导对象,而建议区域则预测一个图像的类标签,而本地化高度相重叠,因此具有非常相近的通量算法,只用于预测边框的RoI集合特性.
<a id="S0367"></a> Source: p.12 S0367
Original: As a result, the image-centric training based on the predicted classes.
中文: 因此,基于预测班次的以形象为中心的训练.
<a id="S0368"></a> Source: p.12 S0368
Original: We adopt the “per-class reof Fast R-CNN [7] generates samples of small variations, gression” (PCR) strategy [40, 41], learning a bounding box which may not be desired for stochastic training.
中文: 我们采用“每级快速R-CNN[7]生成小变异、相向性(PCR)战略[40、41]的样本,学习一个可能不适合进行花样训练的边框。
<a id="S0369"></a> Source: p.12 S0369
Original: We pre-train the networks for Imby this, in our current experiment we use the original RageNet classification and then fine-tune them for localiza- CNN [8] that is RoI-centric, in place of Fast R-CNN. tion.
中文: 我们为Imby预先训练网络, 在目前的实验中, 我们使用原始的RageNet分类, 然后微调它们为localiza-CNN[8], 即以RoI为中心,
<a id="S0370"></a> Source: p.12 S0370
Original: We train networks on the provided 1000-class Ima- Our R-CNN implementation is as follows.
中文: 我们通过提供1 000级的Ima -- -- 我们的R-CNN实施培训网络如下。
<a id="S0371"></a> Source: p.12 S0371
Original: We apply the geNet training set. per-class RPN trained as above on the training images to Our localization algorithm is based on the RPN framepredict bounding boxes for the ground truth class.
中文: 我们使用GeNet的训练集。 每班RPN在训练图像上接受上述训练,以我们本地化算法为基础,采用RPN框架定界框,用于地面真理课.
<a id="S0372"></a> Source: p.12 S0372
Original: These work of [32] with a few modifications.
中文: 这些作品[32]经过了几处修改.
<a id="S0373"></a> Source: p.12 S0373
Original: Unlike the way in predicted boxes play a role of class-dependent proposals. [32] that is category-agnostic, our RPN for localization is For each training image, the highest scored 200 proposals designed in a per-class form.
中文: 与预测框中发挥依赖类建议的作用的方式不同. [32],即类不可知性,我们本地化的RPN是每个训练图像,是每类形式设计得分最高的200个建议.
<a id="S0374"></a> Source: p.12 S0374
Original: This RPN ends with two sibare extracted as training samples to train an R-CNN classiling 1 1 convolutional layers for binary classification (cls) × fier.
中文: 这个RPN结尾取出两个sibare作为训练样品来训练R-CNN分级 1 1 分层为二进制分级(cls)×fier.
<a id="S0375"></a> Source: p.12 S0375
Original: The image region is cropped from a proposal, warped and box regression (reg), as in [32].
中文: 图像区域来自一个提案,被扭曲和框式回归(reg),如[32]。
<a id="S0376"></a> Source: p.12 S0376
Original: The cls and reg layers to 224 224 pixels, and fed into the classification network are both in a per-class from, in contrast to [32].
中文: cls和reg分层为224,224像素,并被输入到分类网络中,两者都是从一个分级中分出,与[32]相对.
<a id="S0377"></a> Source: p.12 S0377
Original: The outputs of this network consist of two cally, the cls layer has a 1000-d output, and each dimension sibling fc layers for cls and reg, also in a per-class form. is binary logistic regression for predicting being or not be- This R-CNN network is fine-tuned on the training set using an object class; the reg layer has a 1000 4-d output × ing a mini-batch size of 256 in the RoI-centric fashion.
中文: 这个网络的输出由两个cally组成,cls层有1000-d的输出,每个维相间的fc层用于cls和reg,也以每等形式. 是二进制逻辑回归 预测是否是... 这个R-CNN网络在使用对象类的训练集上进行了微调;reg层有一个1000个4-d输出×以以RoI为中心时态的256个小批量大小.
<a id="S0378"></a> Source: p.12 S0378
Original: For consisting of box regressors for 1000 classes.
中文: 包括1000个班级的箱式后退器。
<a id="S0379"></a> Source: p.12 S0379
Original: As in [32], testing, the RPN generates the highest scored 200 proposals our bounding box regression is with reference to multiple for each predicted class, and the R-CNN network is used to translation-invariant “anchor” boxes at each position. update these proposals’ scores and box positions.
中文: 如同[32]中,测试,RPN生成最高的得分的200个提案,我们的边框回归是指每个预测类的多个,R-CNN网络用于每个位置的翻译-不变量"锚"框. 更新这些提案的分数和方框位置。
<a id="S0380"></a> Source: p.12 S0380
Original: As in our ImageNet classification training (Sec. 3.4), we This method reduces the top-5 localization error to randomly sample 224 224 crops for data augmentation. × 10.6% (Table 13).
中文: 如同我们的ImageNet分类训练(Sec.3.4)一样,我们这种方法将上到5位本地化误差缩小为随机抽样224,224个作物进行数据增强. × 10.6%(表13)。
<a id="S0381"></a> Source: p.12 S0381
Original: This is our single-model result on the We use a mini-batch size of 256 images for fine-tuning.
中文: 这是我们用256个图像进行微调的小型小批量模型。
<a id="S0382"></a> Source: p.12 S0382
Original: Using an ensemble of networks for both clasavoid negative samples being dominate, 8 anchors are ransification and localization, we achieve a top-5 localization domly sampled for each image, where the sampled positive error of 9.0% on the test set.
中文: 使用一个组合的网络 以Clasavoid阴性标本为主, 8个锚被洗涤 和本地化, 我们实现了一个顶端 5 本地化 多姆抽样 对于每个图像,
<a id="S0383"></a> Source: p.12 S0383
Original: This number significantly outand negative anchors have a ratio of 1:1 [32].
中文: 这一数字显著出差和负锚的比例为1:1 [32].
<a id="S0384"></a> Source: p.12 S0384
Original: For testing, performs the ILSVRC 14 results (Table 14), showing a 64% the network is applied on the image fully-convolutionally. relative reduction of error.
中文: 测试时,执行ILSVRC 14结果(表14),显示网络在图像上完全演化应用了64%. 相对减少出错。
<a id="S0385"></a> Source: p.12 S0385
Original: This result won the 1st place in Table 13 compares the localization results.
中文: 这一结果在表13中获得了第1名,比较了本地化结果.
<a id="S0386"></a> Source: p.12 S0386
Original: Following the ImageNet localization task in ILSVRC 2015. [41], we first perform “oracle” testing using the ground truth class as the classification prediction.
中文: 在ILSVRC2015年的ImageNet本地化任务后[41],我们首先以地真类作为分类预测进行"甲骨文"测试.