This is the ResNet paper. Not only this proposed the shortcut connection architecture, but also give the architecture for image classification and object detection.

This paper answers the question of why deeper network does not always work better. It is not about overfitting, because higher training error is observed. It is because a deeper network is more difficult to train. The key idea is to use residual learning building block, illustrated in figure 2 in the paper. Given a layer learns a function $$F(x)$$, and with shortcut connection effectively the output of the block will be $$H(x) = F(x)+x$$. The paper argues that it is easier to learn the residual $$F(x)$$ than the unreferenced mapping $$H(x)$$.

Specifically, the building block is

\begin{aligned} y &= \sigma( F(x) + x ) \\ &= \sigma( W_2 \cdot \sigma( W_1 x ) + x ) \end{aligned}

where $$\sigma()$$ is the activation function (usually ReLU) and $$W_i$$ is the kernel of the layer (bias is omitted for simplicity). The shortcut connection is usually skipping two layers. But if the depth is different, we may need to transform by a linear projection before add:

$y = \sigma( W_2 \cdot \sigma( W_1 x ) + W_s x )$

ImageNet classification network

The paper constructed a ResNet-34 network with 34 weighted layers. The design is inspired by VGG: mostly conv layers are using 3×3 filters, and for the same output spatial size, the layers have the same output depth but if the output spatial size is halved, the number of filter on the layer is doubled to preserve the time complexity.

VGG-19 operates in 19.6 GFLOPs (multiply-adds), but ResNet-34 only needs 3.6 GFLOPs.

The ResNet-34 model is trained with ImageNet, with 224×224 random crop on the input image with random horizontal flip and pixels are mean-subtracted. And batch normalization is applied after each convolution and before activation. The model is trained with SGD with mini-batch size of 256, for 600K steps. Learning rate of starts at 0.1 and divided by 10 whenever error plateaus. Weight decay of 1e-4 and momentum 0.9 are used, but no dropout is used. Evaluation is using 10-crop testing according to the AlexNet paper.

The paper provided alternative model designs, namely, ResNet-18, ResNet-34, ResNet-50, ResNet-101, and ResNet-152. All of them has 5 stages, with 2 to 36 residual blocks in each stage, and each residual blocks has 2 to 3 convolutional layers. These architectures uses 1.8 GFLOPs to 11.3 GFLOPs.

It is found that with shortcuts, the network converge faster even if the result are comparable (between ResNet-18 with and without shortcut connections). The paper advocates for a parameter-free identity shortcuts (i.e., no $$W_s$$ projection) to reduce computation.

In ResNet-50 to ResNet-152, the residual blocks are with 3 conv layers, and ResNet-18 to ResNet-34 are with 2 conv layers. The 2 conv layer block are two 3×3×N layers with ReLU only on the first layer. The 3 conv layer block, however, is 1×1×N - ReLU - 3×3×N - ReLU - 1×1×4N with the number of channels cut to a quarter at the first conv layer. This is called the bottleneck architecture.

In section 4.2 of the paper, there are also ResNet of 20 to 1202 layers proposed for 32×32 images of CIFAR-10 dataset.

PASCAL and MS COCO detection

Section 4.3 and appendix of the paper proposed to modify Faster R-CNN to replace VGG-16 with ResNet-101. It is found to have 6% increase in COCO’s mAP@[.5, .95] metric, or 28% relative improvement by merely changing this backbone.

According to the appendix, the implementation is as follows:

• ResNet has no hidden FC layers (VGG-16 has 3 FC layers at the output, hence 2 of the FC layers are hidden). The VGG conv feature maps have total stride of 16 pixels w.r.t. input image, and ResNet up to the 4th block also has stride of 16. Thus the feature map right before the 5th block is shared between a RPN and the Fast R-CNN detection network.
• The RPN generates 300 proposals
• On the Fast R-CNN side, RoI pooling layer is applied to the output of 4th block, then the RoI pooled feature map is processed by the 5th block of ResNet, playing the role of the two hidden FC layers in VGG-16, then the output is sent to the two sibling classification and box regression layer
• Batch normalization statistics for each layer is collected on ImageNet training set during pretraining. Then they are fixed during fine-tuning for object detection

Object detection is evaluated using PASCAL VOC metric (mAP@IoU=0.5) and COCO metric (mAP@IoU=.5:.05:.95), using 80K/40K images on the training/validation set. The RPN step has mini-batch size of 8 images and Fast R-CNN step has bathc size of 16 images. The training runs for 240K iterations with learning rate of 1e-3 and another 80K iterations with learning rate of 1e-4.

At inference, the model first take a proposed box and obtain the regressed box from Faster R-CNN output. Then this regressed box runs through the model again to obtain a new classification and new regressed box. With 300 RoI proposed by the RPN, we obtained 300 first step predictions and another 300 second step predictions. Non-maximum suppression is applied on these 600 predictions with IoU threshold of 0.3 and box voting. This two-step process is found to improve mAP by 2 points.

Bibliographic data

@inproceedings{
title = "Deep Residual Learning for Image Recognition",
author = "Kaiming He and Xiangyu Zhang and Shaoqing Ren and Jian Sun",
booktitle = "Proc CVPR",
pages = "770--778",
year = "2016",
note = "arXiv:1512.03385",
}