TTiti的学习笔记
首页 / 专业知识 / 40-References/Papers/resnet - ResNet/01_original.md

Deep Residual Learning for Image Recognition Kaiming He Xiangyu Zhang Shaoqing Ren Jian Sun Microsoft Research kahe, v-xiangz, v-shren, jiansun @microsoft.com

专业知识 · 40-References/Papers/resnet - ResNet/01_original.md

Deep Residual Learning for Image Recognition Kaiming He Xiangyu Zhang Shaoqing Ren Jian Sun Microsoft Research kahe, v-xiangz, v-shren, jiansun @microsoft.com

原文全文

Page 1

<a id="S0001"></a> Source: p.1 S0001

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.

<a id="S0002"></a> Source: p.1 S0002

We present a residual learning framework to ease the training 10 of networks that are substantially deeper than those used previously.

<a id="S0003"></a> Source: p.1 S0003

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.

<a id="S0004"></a> Source: p.1 S0004

We provide comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth.

<a id="S0005"></a> Source: p.1 S0005

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.

<a id="S0006"></a> Source: p.1 S0006

An ensemble of these residual nets achieves 3.57% error on the ImageNet test set.

<a id="S0007"></a> Source: p.1 S0007

This result won the 1st place on the ILSVRC 2015 classification task.

<a id="S0008"></a> Source: p.1 S0008

We also present analysis on CIFAR-10 with 100 and 1000 layers.

<a id="S0009"></a> Source: p.1 S0009

The depth of representations is of central importance for many visual recognition tasks.

<a id="S0010"></a> Source: p.1 S0010

Solely due to our extremely deep representations, we obtain a 28% relative improvement on the COCO object detection dataset.

<a id="S0011"></a> Source: p.1 S0011

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.

<a id="S0012"></a> Source: p.1 S0012

Introduction Deep convolutional neural networks [22, 21] have led to a series of breakthroughs for image classification [21, 50, 40].

<a id="S0013"></a> Source: p.1 S0013

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).

<a id="S0014"></a> Source: p.1 S0014

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].

<a id="S0015"></a> Source: p.1 S0015

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.

<a id="S0016"></a> Source: p.1 S0016

Training error (left) and test error (right) on CIFAR-10 with 20-layer and 56-layer “plain” networks.

<a id="S0017"></a> Source: p.1 S0017

The deeper network has higher training error, and thus test error.

<a id="S0018"></a> Source: p.1 S0018

Similar phenomena on ImageNet is presented in Fig. 4. greatly benefited from very deep models.

<a id="S0019"></a> Source: p.1 S0019

Driven by the significance of depth, a question arises: Is learning better networks as easy as stacking more layers?

<a id="S0020"></a> Source: p.1 S0020

An obstacle to answering this question was the notorious problem of vanishing/exploding gradients [1, 9], which hamper convergence from the beginning.

<a id="S0021"></a> Source: p.1 S0021

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].

<a id="S0022"></a> Source: p.1 S0022

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.

<a id="S0023"></a> Source: p.1 S0023

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.

<a id="S0024"></a> Source: p.1 S0024

The degradation (of training accuracy) indicates that not all systems are similarly easy to optimize.

<a id="S0025"></a> Source: p.1 S0025

Let us consider a shallower architecture and its deeper counterpart that adds more layers onto it.

<a id="S0026"></a> Source: p.1 S0026

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.

<a id="S0027"></a> Source: p.1 S0027

The existence of this constructed solution indicates that a deeper model should produce no higher training error than its shallower counterpart.

<a id="S0028"></a> Source: p.1 S0028

But experiments show that our current solvers on hand are unable to find solutions that 1 5102 ceD 01 ]VC.sc[ 1v58330.2151:viXra

Page 2

<a id="S0029"></a> Source: p.2 S0029

ImageNet test set, and won the 1st place in the ILSVRC x 2015 classification competition.

<a id="S0030"></a> Source: p.2 S0030

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.

<a id="S0031"></a> Source: p.2 S0031

This strong evidence shows that Figure 2.

<a id="S0032"></a> Source: p.2 S0032

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.

<a id="S0033"></a> Source: p.2 S0033

Related Work In this paper, we address the degradation problem by introducing a deep residual learning framework.

<a id="S0034"></a> Source: p.2 S0034

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.

<a id="S0035"></a> Source: p.2 S0035

Formally, denoting the desired formulated as a probabilistic version [18] of VLAD.

<a id="S0036"></a> Source: p.2 S0036

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.

<a id="S0037"></a> Source: p.2 S0037

The orig- trieval and classification [4, 48].

<a id="S0038"></a> Source: p.2 S0038

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.

<a id="S0039"></a> Source: p.2 S0039

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.

<a id="S0040"></a> Source: p.2 S0040

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.

<a id="S0041"></a> Source: p.2 S0041

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).

<a id="S0042"></a> Source: p.2 S0042

These methods suggest that a good reformulation cut connections add neither extra parameter nor computaor preconditioning can simplify the optimization. tional complexity.

<a id="S0043"></a> Source: p.2 S0043

The entire network can still be trained end-to-end by SGD with backpropagation, and can be eas- Shortcut Connections.

<a id="S0044"></a> Source: p.2 S0044

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.

<a id="S0045"></a> Source: p.2 S0045

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].

<a id="S0046"></a> Source: p.2 S0046

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.

<a id="S0047"></a> Source: p.2 S0047

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.

<a id="S0048"></a> Source: p.2 S0048

In [44], an “inception” layer is comsults substantially better than previous networks. posed of a shortcut branch and a few deeper branches.

<a id="S0049"></a> Source: p.2 S0049

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.

<a id="S0050"></a> Source: p.2 S0050

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.

<a id="S0051"></a> Source: p.2 S0051

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.

<a id="S0052"></a> Source: p.2 S0052

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].

<a id="S0053"></a> Source: p.2 S0053

Our ensemble has 3.57% top-5 error on the tional residual functions to be learned.

Page 3

<a id="S0054"></a> Source: p.3 S0054

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.

<a id="S0055"></a> Source: p.3 S0055

The operation + x is performed by a shortcut F connection and element-wise addition.

<a id="S0056"></a> Source: p.3 S0056

Deep Residual Learning ond nonlinearity after the addition (i.e., σ(y), see Fig. 2).

<a id="S0057"></a> Source: p.3 S0057

The shortcut connections in Eqn.(1) introduce neither ex- 3.1.

<a id="S0058"></a> Source: p.3 S0058

Residual Learning tra parameter nor computation complexity.

<a id="S0059"></a> Source: p.3 S0059

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.

<a id="S0060"></a> Source: p.3 S0060

We can fairly comwith x denoting the inputs to the first of these layers.

<a id="S0061"></a> Source: p.3 S0061

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).

<a id="S0062"></a> Source: p.3 S0062

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.

<a id="S0063"></a> Source: p.3 S0063

The original function thus becomes y = (x, W i ) + W s x. (2) F H − F { } (x)+x.

<a id="S0064"></a> Source: p.3 S0064

Although both forms should be able to asymptot- F We can also use a square matrix W in Eqn.(1).

<a id="S0065"></a> Source: p.3 S0065

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).

<a id="S0066"></a> Source: p.3 S0066

As s The form of the residual function is flexible.

<a id="S0067"></a> Source: p.3 S0067

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.

<a id="S0068"></a> Source: p.3 S0068

But if have training error no greater than its shallower counterhas only a single layer, Eqn.(1) is similar to a linear layer: part.

<a id="S0069"></a> Source: p.3 S0069

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.

<a id="S0070"></a> Source: p.3 S0070

With the residual learning refully-connected layers for simplicity, they are applicable to formulation, if identity mappings are optimal, the solvers convolutional layers.

<a id="S0071"></a> Source: p.3 S0071

The function (x, W ) can repremay simply drive the weights of the multiple nonlinear lay- i F { } sent multiple convolutional layers.

<a id="S0072"></a> Source: p.3 S0072

The element-wise addiers toward zero to approach identity mappings. tion is performed on two feature maps, channel by channel.

<a id="S0073"></a> Source: p.3 S0073

In real cases, it is unlikely that identity mappings are optimal, but our reformulation may help to precondition the 3.3.

<a id="S0074"></a> Source: p.3 S0074

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.

<a id="S0075"></a> Source: p.3 S0075

To provide instances for dismapping, than to learn the function as a new one.

<a id="S0076"></a> Source: p.3 S0076

We show cussion, we describe two models for ImageNet as follows. by experiments (Fig. 7) that the learned residual functions in Plain Network.

<a id="S0077"></a> Source: p.3 S0077

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).

<a id="S0078"></a> Source: p.3 S0078

The convolutional layers mostly have 3 3 filters and × 3.2.

<a id="S0079"></a> Source: p.3 S0079

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.

<a id="S0080"></a> Source: p.3 S0080

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.

<a id="S0081"></a> Source: p.3 S0081

We perform downsampling directly by convolutional layers that have a stride of 2.

<a id="S0082"></a> Source: p.3 S0082

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.

<a id="S0083"></a> Source: p.3 S0083

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.

<a id="S0084"></a> Source: p.3 S0084

The function (x, W ) represents the F { i } It is worth noticing that our model has fewer filters and residual mapping to be learned.

<a id="S0085"></a> Source: p.3 S0085

For the example in Fig. 2 lower complexity than VGG nets [41] (Fig. 3, left).

<a id="S0086"></a> Source: p.3 S0086

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.

<a id="S0087"></a> Source: p.3 S0087

See [28]. is only 18% of VGG-19 (19.6 billion FLOPs). 3

Page 4

<a id="S0088"></a> Source: p.4 S0088

VGG-19 34-layer plain 34-layer residual Residual Network.

<a id="S0089"></a> Source: p.4 S0089

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.

<a id="S0090"></a> Source: p.4 S0090

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).

<a id="S0091"></a> Source: p.4 S0091

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.

<a id="S0092"></a> Source: p.4 S0092

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).

<a id="S0093"></a> Source: p.4 S0093

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.

<a id="S0094"></a> Source: p.4 S0094

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].

<a id="S0095"></a> Source: p.4 S0095

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].

<a id="S0096"></a> Source: p.4 S0096

The 3x3 conv, 512 3x3 conv, 128 3x3 conv, 128 standard color augmentation in [21] is used.

<a id="S0097"></a> Source: p.4 S0097

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].

<a id="S0098"></a> Source: p.4 S0098

We initialize the weights 3x3 conv, 128 3x3 conv, 128 as in [13] and train all plain/residual nets from scratch.

<a id="S0099"></a> Source: p.4 S0099

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.

<a id="S0100"></a> Source: p.4 S0100

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.

<a id="S0101"></a> Source: p.4 S0101

We 3x3 conv, 512 3x3 conv, 256 3x3 conv, 256 × use a weight decay of 0.0001 and a momentum of 0.9.

<a id="S0102"></a> Source: p.4 S0102

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].

<a id="S0103"></a> Source: p.4 S0103

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.

<a id="S0104"></a> Source: p.4 S0104

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.

<a id="S0105"></a> Source: p.4 S0105

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.

<a id="S0106"></a> Source: p.4 S0106

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.

<a id="S0107"></a> Source: p.4 S0107

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.

<a id="S0108"></a> Source: p.4 S0108

We first evaluate 18-layer and 34-layer plain nets.

<a id="S0109"></a> Source: p.4 S0109

The 34-layer plain net is in Fig. 3 (middle).

<a id="S0110"></a> Source: p.4 S0110

Example network architectures for ImageNet.

<a id="S0111"></a> Source: p.4 S0111

Left: the 18-layer plain net is of a similar form.

<a id="S0112"></a> Source: p.4 S0112

See Table 1 for de- VGG-19 model [41] (19.6 billion FLOPs) as a reference.

<a id="S0113"></a> Source: p.4 S0113

Mid- tailed architectures. dle: a plain network with 34 parameter layers (3.6 billion FLOPs).

<a id="S0114"></a> Source: p.4 S0114

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).

<a id="S0115"></a> Source: p.4 S0115

The dotted shortcuts increase dimensions.

<a id="S0116"></a> Source: p.4 S0116

To reveal the reasons, in Fig. 4 (left) we commore details and other variants. pare their training/validation errors during the training procedure.

<a id="S0117"></a> Source: p.4 S0117

We have observed the degradation problem - the 4

Page 5

<a id="S0118"></a> Source: p.5 S0118

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.

<a id="S0119"></a> Source: p.5 S0119

Building blocks are shown in brackets (see also Fig. 5), with the numbers of blocks stacked.

<a id="S0120"></a> Source: p.5 S0120

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.

<a id="S0121"></a> Source: p.5 S0121

Thin curves denote training error, and bold curves denote validation error of the center crops.

<a id="S0122"></a> Source: p.5 S0122

Left: plain networks of 18 and 34 layers.

<a id="S0123"></a> Source: p.5 S0123

In this plot, the residual networks have no extra parameter compared to their plain counterparts. plain ResNet reducing of the training error3.

<a id="S0124"></a> Source: p.5 S0124

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.

<a id="S0125"></a> Source: p.5 S0125

Next we evaluate 18-layer and 34- Table 2.

<a id="S0126"></a> Source: p.5 S0126

Top-1 error (%, 10-crop testing) on ImageNet validation. layer residual nets (ResNets).

<a id="S0127"></a> Source: p.5 S0127

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.

<a id="S0128"></a> Source: p.5 S0128

Fig. 4 shows the training procedures. connection is added to each pair of 3 3 filters as in Fig. 3 × (right).

<a id="S0129"></a> Source: p.5 S0129

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).

<a id="S0130"></a> Source: p.5 S0130

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.

<a id="S0131"></a> Source: p.5 S0131

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.

<a id="S0132"></a> Source: p.5 S0132

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.

<a id="S0133"></a> Source: p.5 S0133

We also verify that the validation data.

<a id="S0134"></a> Source: p.5 S0134

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.

<a id="S0135"></a> Source: p.5 S0135

So neither forward nor backward signals vanish.

<a id="S0136"></a> Source: p.5 S0136

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.

<a id="S0137"></a> Source: p.5 S0137

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

Page 6

<a id="S0138"></a> Source: p.6 S0138

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.

<a id="S0139"></a> Source: p.6 S0139

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.

<a id="S0140"></a> Source: p.6 S0140

Right: a “bottleneck” building block for ResNet-50/101/152.

<a id="S0141"></a> Source: p.6 S0141

ResNet-101 21.75 6.05 ResNet-152 21.43 5.71 parameter-free, identity shortcuts help with training.

<a id="S0142"></a> Source: p.6 S0142

Error rates (%, 10-crop testing) on ImageNet validation. we investigate projection shortcuts (Eqn.(2)).

<a id="S0143"></a> Source: p.6 S0143

In Table 3 we VGG-16 is based on our test.

<a id="S0144"></a> Source: p.6 S0144

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.

<a id="S0145"></a> Source: p.6 S0145

VGG [41] (ILSVRC’14) - 8.43† tion shortcuts are used for increasing dimensions, and other shortcuts are identity; and (C) all shortcuts are projections.

<a id="S0146"></a> Source: p.6 S0146

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.

<a id="S0147"></a> Source: p.6 S0147

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.

<a id="S0148"></a> Source: p.6 S0148

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.

<a id="S0149"></a> Source: p.6 S0149

So we ResNet-152 19.38 4.49 do not use option C in the rest of this paper, to reduce mem- Table 4.

<a id="S0150"></a> Source: p.6 S0150

Error rates (%) of single-model results on the ImageNet ory/time complexity and model sizes.

<a id="S0151"></a> Source: p.6 S0151

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.

<a id="S0152"></a> Source: p.6 S0152

Next we describe our GoogLeNet [44] (ILSVRC’14) 6.66 deeper nets for ImageNet.

<a id="S0153"></a> Source: p.6 S0153

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.

<a id="S0154"></a> Source: p.6 S0154

For each residual function , we F BN-inception [16] 4.82 use a stack of 3 layers instead of 2 (Fig. 5).

<a id="S0155"></a> Source: p.6 S0155

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.

<a id="S0156"></a> Source: p.6 S0156

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.

<a id="S0157"></a> Source: p.6 S0157

Fig. 5 shows an example, where both designs have similar time complexity.

<a id="S0158"></a> Source: p.6 S0158

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.

<a id="S0159"></a> Source: p.6 S0159

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.

<a id="S0160"></a> Source: p.6 S0160

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).

<a id="S0161"></a> Source: p.6 S0161

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.

<a id="S0162"></a> Source: p.6 S0162

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.

<a id="S0163"></a> Source: p.6 S0163

So the usage of bottleneck designs is mainly due to practical considerations.

<a id="S0164"></a> Source: p.6 S0164

We further note that the degradation problem Identity vs.

<a id="S0165"></a> Source: p.6 S0165

We have shown that of plain nets is also witnessed for the bottleneck designs. 6

Page 7

<a id="S0166"></a> Source: p.7 S0166

34-layer net with this 3-layer bottleneck block, resulting in method error (%) a 50-layer ResNet (Table 1).

<a id="S0167"></a> Source: p.7 S0167

We use option B for increasing Maxout [10] 9.38 dimensions.

<a id="S0168"></a> Source: p.7 S0168

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).

<a id="S0169"></a> Source: p.7 S0169

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).

<a id="S0170"></a> Source: p.7 S0170

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).

<a id="S0171"></a> Source: p.7 S0171

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.

<a id="S0172"></a> Source: p.7 S0172

The benefits of depth are witnessed for all evaluation ResNet 1202 19.4M 7.93 metrics (Table 3 and 4).

<a id="S0173"></a> Source: p.7 S0173

Classification error on the CIFAR-10 test set.

<a id="S0174"></a> Source: p.7 S0174

All meth- Comparisons with State-of-the-art Methods.

<a id="S0175"></a> Source: p.7 S0175

In Table 4 ods are with data augmentation.

<a id="S0176"></a> Source: p.7 S0176

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].

<a id="S0177"></a> Source: p.7 S0177

Our baseline 34-layer ResNets have achieved very competitive accuracy.

<a id="S0178"></a> Source: p.7 S0178

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%.

<a id="S0179"></a> Source: p.7 S0179

This single-model result and number of parameters as the plain counterparts. outperforms all previous ensemble results (Table 5).

<a id="S0180"></a> Source: p.7 S0180

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.

<a id="S0181"></a> Source: p.7 S0181

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.

<a id="S0182"></a> Source: p.7 S0182

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.

<a id="S0183"></a> Source: p.7 S0183

CIFAR-10 and Analysis terminate training at 64k iterations, which is determined on a 45k/5k train/val split.

<a id="S0184"></a> Source: p.7 S0184

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.

<a id="S0185"></a> Source: p.7 S0185

We present experiments trained × image or its horizontal flip.

<a id="S0186"></a> Source: p.7 S0186

For testing, we only evaluate on the training set and evaluated on the test set.

<a id="S0187"></a> Source: p.7 S0187

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.

<a id="S0188"></a> Source: p.7 S0188

Fig. 6 (left) shows the behaviors of the simple architectures as follows. plain nets.

<a id="S0189"></a> Source: p.7 S0189

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.

<a id="S0190"></a> Source: p.7 S0190

The network inputs are 32 32 images, with phenomenon is similar to that on ImageNet (Fig. 4, left) and × the per-pixel mean subtracted.

<a id="S0191"></a> Source: p.7 S0191

The first layer is 3 3 convoon MNIST (see [42]), suggesting that such an optimization × lutions.

<a id="S0192"></a> Source: p.7 S0192

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.

<a id="S0193"></a> Source: p.7 S0193

Also { } with 2n layers for each feature map size.

<a id="S0194"></a> Source: p.7 S0194

The numbers of similar to the ImageNet cases (Fig. 4, right), our ResNets filters are 16, 32, 64 respectively.

<a id="S0195"></a> Source: p.7 S0195

The subsampling is permanage to overcome the optimization difficulty and demon- { } formed by convolutions with a stride of 2.

<a id="S0196"></a> Source: p.7 S0196

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.

<a id="S0197"></a> Source: p.7 S0197

There are totally 6n+2 stacked weighted ResNet.

<a id="S0198"></a> Source: p.7 S0198

In this case, we find that the initial learning rate layers.

<a id="S0199"></a> Source: p.7 S0199

The following table summarizes the architecture: of 0.1 is slightly too large to start converging5.

<a id="S0200"></a> Source: p.7 S0200

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.

<a id="S0201"></a> Source: p.7 S0201

The rest of the learning schedule is as done # filters 16 32 64 previously.

<a id="S0202"></a> Source: p.7 S0202

This 110-layer network converges well (Fig. 6, middle).

<a id="S0203"></a> Source: p.7 S0203

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).

<a id="S0204"></a> Source: p.7 S0204

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

Page 8

<a id="S0205"></a> Source: p.8 S0205

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.

<a id="S0206"></a> Source: p.8 S0206

Dashed lines denote training error, and bold lines denote testing error.

<a id="S0207"></a> Source: p.8 S0207

The error of plain-110 is higher than 60% and not displayed.

<a id="S0208"></a> Source: p.8 S0208

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.

<a id="S0209"></a> Source: p.8 S0209

Object detection mAP (%) on the PASCAL VOC 2007/2012 test sets using baseline Faster R-CNN.

<a id="S0210"></a> Source: p.8 S0210

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.

<a id="S0211"></a> Source: p.8 S0211

Standard deviations (std) of layer responses on CIFAR- 10.

<a id="S0212"></a> Source: p.8 S0212

The responses are the outputs of each 3×3 layer, after BN and Table 8.

<a id="S0213"></a> Source: p.8 S0213

Object detection mAP (%) on the COCO validation set before nonlinearity.

<a id="S0214"></a> Source: p.8 S0214

Top: the layers are shown in their original using baseline Faster R-CNN.

<a id="S0215"></a> Source: p.8 S0215

See also Table 9 for better results. order.

<a id="S0216"></a> Source: p.8 S0216

Bottom: the responses are ranked in descending order. have similar training error.

<a id="S0217"></a> Source: p.8 S0217

We argue that this is because of networks such as FitNet [35] and Highway [42] (Table 6), overfitting.

<a id="S0218"></a> Source: p.8 S0218

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.

<a id="S0219"></a> Source: p.8 S0219

Strong regularization such as maxout [10] or dropout [14] is applied to obtain the Analysis of Layer Responses.

<a id="S0220"></a> Source: p.8 S0220

Fig. 7 shows the standard best results ([10, 25, 24, 35]) on this dataset.

<a id="S0221"></a> Source: p.8 S0221

In this paper, deviations (std) of the layer responses.

<a id="S0222"></a> Source: p.8 S0222

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).

<a id="S0223"></a> Source: p.8 S0223

For ResNets, this analydistracting from the focus on the difficulties of optimizasis reveals the response strength of the residual functions. tion.

<a id="S0224"></a> Source: p.8 S0224

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.

<a id="S0225"></a> Source: p.8 S0225

These results support our basic motivation (Sec.3.1) that the residual functions might 4.3.

<a id="S0226"></a> Source: p.8 S0226

Object Detection on PASCAL and MS COCO be generally closer to zero than the non-residual functions.

<a id="S0227"></a> Source: p.8 S0227

Our method has good generalization performance on We also notice that the deeper ResNet has smaller magniother recognition tasks.

<a id="S0228"></a> Source: p.8 S0228

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.

<a id="S0229"></a> Source: p.8 S0229

We adopt Faster R-CNN [32] as the delayers, an individual layer of ResNets tends to modify the tection method.

<a id="S0230"></a> Source: p.8 S0230

Here we are interested in the improvements signal less. of replacing VGG-16 [41] with ResNet-101.

<a id="S0231"></a> Source: p.8 S0231

The detection Exploring Over 1000 layers.

<a id="S0232"></a> Source: p.8 S0232

We explore an aggressively implementation (see appendix) of using both models is the deep model of over 1000 layers.

<a id="S0233"></a> Source: p.8 S0233

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.

<a id="S0234"></a> Source: p.8 S0234

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.

<a id="S0235"></a> Source: p.8 S0235

Its test error is still fairly good solely due to the learned representations. (7.93%, Table 6).

<a id="S0236"></a> Source: p.8 S0236

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.

<a id="S0237"></a> Source: p.8 S0237

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.

Page 9

<a id="S0238"></a> Source: p.9 S0238

On the number of linear regions of deep neural networks.

<a id="S0239"></a> Source: p.9 S0239

Rectified linear units improve restricted cies with gradient descent is difficult.

<a id="S0240"></a> Source: p.9 S0240

IEEE Transactions on Neural boltzmann machines.

<a id="S0241"></a> Source: p.9 S0241

Fisher kernels on visual vocabularies for [2] C. M.

<a id="S0242"></a> Source: p.9 S0242

Neural networks for pattern recognition.

<a id="S0243"></a> Source: p.9 S0243

In CVPR, 2007. university press, 1995. [31] T.

<a id="S0244"></a> Source: p.9 S0244

Siam, linear transformations in perceptrons.

<a id="S0245"></a> Source: p.9 S0245

The devil real-time object detection with region proposal networks.

<a id="S0246"></a> Source: p.9 S0246

In NIPS, is in the details: an evaluation of recent feature encoding methods. 2015.

<a id="S0247"></a> Source: p.9 S0247

Zisnetworks on convolutional feature maps. arXiv:1504.06066, 2015. serman.

<a id="S0248"></a> Source: p.9 S0248

The Pascal Visual Object Classes (VOC) Challenge.

<a id="S0249"></a> Source: p.9 S0249

Pattern recognition and neural networks.

<a id="S0250"></a> Source: p.9 S0250

Cambridge pages 303–338, 2010. university press, 1996. [6] S.

<a id="S0251"></a> Source: p.9 S0251

Object detection via a multi-region & [35] A.

<a id="S0252"></a> Source: p.9 S0252

Gatta, and semantic segmentation-aware cnn model.

<a id="S0253"></a> Source: p.9 S0253

Imagenet archies for accurate object detection and semantic segmentation.

<a id="S0254"></a> Source: p.9 S0254

In large scale visual recognition challenge. arXiv:1409.0575, 2014.

<a id="S0255"></a> Source: p.9 S0255

Understanding the difficulty of training the nonlinear dynamics of learning in deep linear neural networks. deep feedforward neural networks.

<a id="S0256"></a> Source: p.9 S0256

In AISTATS, 2010. arXiv:1312.6120, 2013. [10] I. J.

<a id="S0257"></a> Source: p.9 S0257

Accelerated gradient descent by factor-centering Y.

<a id="S0258"></a> Source: p.9 S0258

Maxout networks. arXiv:1302.4389, 2013. decomposition.

<a id="S0259"></a> Source: p.9 S0259

Convolutional neural networks at constrained time [39] N. N.

<a id="S0260"></a> Source: p.9 S0260

Centering neural network gradient factors.

<a id="S0261"></a> Source: p.9 S0261

Neural Networks: Tricks of the Trade, pages 207–226.

<a id="S0262"></a> Source: p.9 S0262

Spatial pyramid pooling in deep 1998. convolutional networks for visual recognition.

<a id="S0263"></a> Source: p.9 S0263

Overfeat: Integrated recognition, localization and detection Surpassing human-level performance on imagenet classification.

<a id="S0264"></a> Source: p.9 S0264

Very deep convolutional networks [14] G. E.

<a id="S0265"></a> Source: p.9 S0265

Sutskever, and for large-scale image recognition.

<a id="S0266"></a> Source: p.9 S0266

Improving neural networks by preventing co- [42] R. K.

<a id="S0267"></a> Source: p.9 S0267

Highway networks. adaptation of feature detectors. arXiv:1207.0580, 2012. arXiv:1505.00387, 2015. [15] S.

<a id="S0268"></a> Source: p.9 S0268

Training very deep computation, 9(8):1735–1780, 1997. networks. 1507.06228, 2015. [16] S.

<a id="S0269"></a> Source: p.9 S0269

Batch normalization: Accelerating deep [44] C.

<a id="S0270"></a> Source: p.9 S0270

Ernetwork training by reducing internal covariate shift.

<a id="S0271"></a> Source: p.9 S0271

Fast surface interpolation using hierarchical basis func- [18] H.

<a id="S0272"></a> Source: p.9 S0272

Aggregating local image descriptors into compact codes. [46] R.

<a id="S0273"></a> Source: p.9 S0273

Locally adapted hierarchical basis preconditioning.

<a id="S0274"></a> Source: p.9 S0274

Caffe: Convolutional architecture for tic gradient towards second-order methods–backpropagation learnfast feature embedding. arXiv:1408.5093, 2014. ing with transformations in nonlinearities.

<a id="S0275"></a> Source: p.9 S0275

Learning multiple layers of features from tiny im- Processing, 2013. ages.

<a id="S0276"></a> Source: p.9 S0276

VLFeat: An open and portable library [21] A.

<a id="S0277"></a> Source: p.9 S0277

Imagenet classification of computer vision algorithms, 2008. with deep convolutional neural networks.

<a id="S0278"></a> Source: p.9 S0278

Modern applied statistics with s-plus. [22] Y.

<a id="S0279"></a> Source: p.9 S0279

Backpropagation applied to hand- [50] M. D.

<a id="S0280"></a> Source: p.9 S0280

Visualizing and understanding convoluwritten zip code recognition.

<a id="S0281"></a> Source: p.9 S0281

Neural computation, 1989. tional neural networks.

<a id="S0282"></a> Source: p.9 S0282

In Neural Networks: Tricks of the Trade, pages 9–50.

<a id="S0283"></a> Source: p.9 S0283

Deeplysupervised nets. arXiv:1409.5185, 2014. [25] M.

<a id="S0284"></a> Source: p.9 S0284

Network in network. arXiv:1312.4400, 2013. [26] T.-Y.

<a id="S0285"></a> Source: p.9 S0285

Microsoft COCO: Common objects in context.

<a id="S0286"></a> Source: p.9 S0286

Fully convolutional networks for semantic segmentation.

Page 10

<a id="S0287"></a> Source: p.10 S0287

Object Detection Baselines 8 images (i.e., 1 per GPU) and the Fast R-CNN step has a mini-batch size of 16 images.

<a id="S0288"></a> Source: p.10 S0288

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.

<a id="S0289"></a> Source: p.10 S0289

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.

<a id="S0290"></a> Source: p.10 S0290

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.

<a id="S0291"></a> Source: p.10 S0291

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.

<a id="S0292"></a> Source: p.10 S0292

We adopt the idea of “Networks on Conv feanearly as big as mAP@.5’s (6.9%).

<a id="S0293"></a> Source: p.10 S0293

This suggests that a ture maps” (NoC) [33] to address this issue.

<a id="S0294"></a> Source: p.10 S0294

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.

<a id="S0295"></a> Source: p.10 S0295

Object Detection Improvements layers in ResNet-101; Table 1).

<a id="S0296"></a> Source: p.10 S0296

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.

<a id="S0297"></a> Source: p.10 S0297

These improvements are based on deep same total stride (16 pixels).

<a id="S0298"></a> Source: p.10 S0298

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].

<a id="S0299"></a> Source: p.10 S0299

RoI pool- MS COCO ing [7] is performed before conv5 1.

<a id="S0300"></a> Source: p.10 S0300

Our box refinement partially follows the itfeature, all layers of conv5 x and up are adopted for each erative localization in [6].

<a id="S0301"></a> Source: p.10 S0301

In Faster R-CNN, the final output region, playing the roles of VGG-16’s fc layers.

<a id="S0302"></a> Source: p.10 S0302

The final is a regressed box that is different from its proposal box.

<a id="S0303"></a> Source: p.10 S0303

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.

<a id="S0304"></a> Source: p.10 S0304

We combine these 300 new predictions with the origpute the BN statistics (means and variances) for each layer inal 300 predictions.

<a id="S0305"></a> Source: p.10 S0305

Non-maximum suppression (NMS) is on the ImageNet training set.

<a id="S0306"></a> Source: p.10 S0306

Then the BN layers are fixed applied on the union set of predicted boxes using an IoU during fine-tuning for object detection.

<a id="S0307"></a> Source: p.10 S0307

As such, the BN threshold of 0.3 [8], followed by box voting [6].

<a id="S0308"></a> Source: p.10 S0308

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.

<a id="S0309"></a> Source: p.10 S0309

We combine global context in the Fast fix the BN layers mainly for reducing memory consumption R-CNN step.

<a id="S0310"></a> Source: p.10 S0310

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.

<a id="S0311"></a> Source: p.10 S0311

This pooled feature is fed into the post-RoI layers to val images in VOC 2012 for training (“07+12”).

<a id="S0312"></a> Source: p.10 S0312

For the obtain a global context feature.

<a id="S0313"></a> Source: p.10 S0313

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.

<a id="S0314"></a> Source: p.10 S0314

The hyper-parameters for train- new structure is trained end-to-end.

<a id="S0315"></a> Source: p.10 S0315

Global context iming Faster R-CNN are the same as in [32].

<a id="S0316"></a> Source: p.10 S0316

Table 7 shows proves mAP@.5 by about 1 point (Table 9). the results.

<a id="S0317"></a> Source: p.10 S0317

ResNet-101 improves the mAP by >3% over Multi-scale testing.

<a id="S0318"></a> Source: p.10 S0318

In the above, all results are obtained by VGG-16.

<a id="S0319"></a> Source: p.10 S0319

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.

<a id="S0320"></a> Source: p.10 S0320

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.

<a id="S0321"></a> Source: p.10 S0321

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).

<a id="S0322"></a> Source: p.10 S0322

We use the 80k images on the train set for train- training because of limited time.

<a id="S0323"></a> Source: p.10 S0323

In addition, we have pering and the 40k images on the val set for evaluation.

<a id="S0324"></a> Source: p.10 S0324

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).

<a id="S0325"></a> Source: p.10 S0325

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

Page 11

<a id="S0326"></a> Source: p.11 S0326

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.

<a id="S0327"></a> Source: p.11 S0327

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.

<a id="S0328"></a> Source: p.11 S0328

Detection results on the PASCAL VOC 2007 test set.

<a id="S0329"></a> Source: p.11 S0329

The baseline is the Faster R-CNN system.

<a id="S0330"></a> Source: p.11 S0330

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.

<a id="S0331"></a> Source: p.11 S0331

Detection results on the PASCAL VOC 2012 test set (http://host.robots.ox.ac.uk:8080/leaderboard/ displaylb.php?challengeid=11&compid=4).

<a id="S0332"></a> Source: p.11 S0332

The baseline is the Faster R-CNN system.

<a id="S0333"></a> Source: p.11 S0333

The system “baseline+++” include box refinement, context, and multi-scale testing in Table 9.

<a id="S0334"></a> Source: p.11 S0334

We select two adjacent scales from the pyramid following val2 test [33].

<a id="S0335"></a> Source: p.11 S0335

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].

<a id="S0336"></a> Source: p.11 S0336

Multi-scale testing improves the mAP our ensemble (ILSVRC’15) 63.6 62.1 by over 2 points (Table 9).

<a id="S0337"></a> Source: p.11 S0337

Our results (mAP, %) on the ImageNet detection dataset.

<a id="S0338"></a> Source: p.11 S0338

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.

<a id="S0339"></a> Source: p.11 S0339

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.

<a id="S0340"></a> Source: p.11 S0340

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.

<a id="S0341"></a> Source: p.11 S0341

This is our single-model result. on PASCAL VOC 2012 is 10 points higher than the previ- Ensemble.

<a id="S0342"></a> Source: p.11 S0342

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.

<a id="S0343"></a> Source: p.11 S0343

We use an ensemble for The ImageNet Detection (DET) task involves 200 object proposing regions, and the union set of proposals are procategories.

<a id="S0344"></a> Source: p.11 S0344

Our cessed by an ensemble of per-region classifiers.

<a id="S0345"></a> Source: p.11 S0345

Table 9 object detection algorithm for ImageNet DET is the same shows our result based on an ensemble of 3 networks.

<a id="S0346"></a> Source: p.11 S0346

The networks are premAP is 59.0% and 37.4% on the test-dev set.

<a id="S0347"></a> Source: p.11 S0347

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.

<a id="S0348"></a> Source: p.11 S0348

We split the validation set PASCAL VOC into two parts (val1/val2) following [8].

<a id="S0349"></a> Source: p.11 S0349

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.

<a id="S0350"></a> Source: p.11 S0350

With the single model on the COCO dataset (55.7% set.

<a id="S0351"></a> Source: p.11 S0351

We do not use other mAP@.5 in Table 9), we fine-tune this model on the PAS- ILSVRC 2015 data.

<a id="S0352"></a> Source: p.11 S0352

Our single model with ResNet-101 has CAL VOC sets.

<a id="S0353"></a> Source: p.11 S0353

The improvements of box refinement, con- 6http://host.robots.ox.ac.uk:8080/anonymous/3OJ4OJ.html, text, and multi-scale testing are also adopted.

Page 12

<a id="S0354"></a> Source: p.12 S0354

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.

<a id="S0355"></a> Source: p.12 S0355

Comparisons of localization error (%) on the ImageNet Table 13.

<a id="S0356"></a> Source: p.12 S0356

Localization error (%) on the ImageNet validation.

<a id="S0357"></a> Source: p.12 S0357

In dataset with state-of-the-art methods. the column of “LOC error on GT class” ([41]), the ground truth class is used.

<a id="S0358"></a> Source: p.12 S0358

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.

<a id="S0359"></a> Source: p.12 S0359

Under the same setting, our RPN method using ResNet-101 net significantly reduces the center-crop error to 13.3%.

<a id="S0360"></a> Source: p.12 S0360

This comparison demonstrates the excellent 58.8% mAP and our ensemble of 3 models has 62.1% mAP performance of our framework.

<a id="S0361"></a> Source: p.12 S0361

With dense (fully convoluon the DET test set (Table 12).

<a id="S0362"></a> Source: p.12 S0362

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.

<a id="S0363"></a> Source: p.12 S0363

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.

<a id="S0364"></a> Source: p.12 S0364

ImageNet Localization The above results are only based on the proposal network (RPN) in Faster R-CNN [32].

<a id="S0365"></a> Source: p.12 S0365

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.

<a id="S0366"></a> Source: p.12 S0366

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.

<a id="S0367"></a> Source: p.12 S0367

As a result, the image-centric training based on the predicted classes.

<a id="S0368"></a> Source: p.12 S0368

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.

<a id="S0369"></a> Source: p.12 S0369

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.

<a id="S0370"></a> Source: p.12 S0370

We train networks on the provided 1000-class Ima- Our R-CNN implementation is as follows.

<a id="S0371"></a> Source: p.12 S0371

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.

<a id="S0372"></a> Source: p.12 S0372

These work of [32] with a few modifications.

<a id="S0373"></a> Source: p.12 S0373

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.

<a id="S0374"></a> Source: p.12 S0374

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.

<a id="S0375"></a> Source: p.12 S0375

The image region is cropped from a proposal, warped and box regression (reg), as in [32].

<a id="S0376"></a> Source: p.12 S0376

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].

<a id="S0377"></a> Source: p.12 S0377

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.

<a id="S0378"></a> Source: p.12 S0378

For consisting of box regressors for 1000 classes.

<a id="S0379"></a> Source: p.12 S0379

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.

<a id="S0380"></a> Source: p.12 S0380

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).

<a id="S0381"></a> Source: p.12 S0381

This is our single-model result on the We use a mini-batch size of 256 images for fine-tuning.

<a id="S0382"></a> Source: p.12 S0382

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.

<a id="S0383"></a> Source: p.12 S0383

This number significantly outand negative anchors have a ratio of 1:1 [32].

<a id="S0384"></a> Source: p.12 S0384

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.

<a id="S0385"></a> Source: p.12 S0385

This result won the 1st place in Table 13 compares the localization results.

<a id="S0386"></a> Source: p.12 S0386

Following the ImageNet localization task in ILSVRC 2015. [41], we first perform “oracle” testing using the ground truth class as the classification prediction.