1.Deep High-Resolution Representation Learning for Human Pose
Estimation(HRNetV1)
2.High-Resolution Representations for Labeling Pixels and
Regions(HRNetV2,HRNetV2p)
1.Introduction
Human pose estimation （ Also known as key point detection ） Designed from the size of W*H*3 Image of I Medium detection K A key point or part （ for example , elbow , Wrist, etc ） The location of . The state-of-the-art method transforms this problem into estimation K Size of W*
H Heat map of {H1,H2,…,HK}, Each heat map Hk Denotes the second k Location confidence of key points .
Typical attitude estimation network
Hourglass： Symmetric codec network
Cascaded pyramid networks：refinenet The feature images of different scales are convoluted and fused
SimpleBaseline： Resolution recovery using transposed convolution in decoder
Combination with dilated convolutions： Using hole convolution to enlarge receptive field in encoder
The characteristics of these networks ：
Two processes :High-to-Low process ( Generating low resolution advanced feature representation ) and Low-to-High( Restore high resolution ) process , And the two processes are serial .
Some networks will fuse high-level feature graph with low-level feature graph .
2.network architecture
advantage ：1. High resolution feature representation is maintained throughout the process , Gradually increase High-to-Low Subnet of , The multiresolution subnets are connected in parallel .
2. Exchanging information repeatedly between parallel multiresolution subnets , Multi scale fusion , High resolution features and low resolution features enhance each other .
The network is divided into 4 individual stage, each stage It's better than the last one stage One more branch , The new branch is to the previous one stage All feature maps are processed strided
convolution The results after fusion , The resolution size is half that of the previous branch , Double the number of channels , each stage from mutil-resolution block form .
every last mutil-resolution block It is divided into two parts :
(a) multi-resolution group convolution: Several parallel branches , Each branch contains 4 Residual units
(b) multi-resolution convolution(exchange unit): Multi scale feature fusion is carried out
The third 3 individual stage Of exchange unit Sketch Map ：
high , in , Three low resolution feature maps are fused with each other : Use for high resolution images strided
convolution, The low resolution image is up sampled and processed 1*1 Convolution of , Because the fusion strategy is to add elements , It is necessary to adjust the channel number of different resolution feature maps to the same number .
3. Human posture detection
HRNetV1: Only output high resolution feature map
experimental result ：
experiment ：COCO Keypoint Detection
Experimental results of verification set
And the best at the moment SimpleBaseline compare ,HRNet-W32( Number of channels :32,64,128,256) and HRNet-W48( Number of channels :48,96,192,384) Use fewer parameters , Less computation and higher performance .
4. Semantic segmentation , Face key point detection
HRNetV2: Use all resolution feature maps , After sampling on the low resolution feature map, it is spliced with the high resolution feature map , after 1*1 convolution ,softmax Layer generated segmentation prediction graph
experiment :Cityscapes Segmentation
Validation set ：
HRNetV2-W40 In comparison UNet++,DeepLabv3 When the number of parameters is less , Achieve higher quality mIOU
HRNetV2-W48 In and PSPNet The parameters are the same , Achieve higher quality mIOU
Test set ：
experiment ：PASCAL context Segmentation
experimental result
In the evaluation method without background tag and with background tag ,HRNetV2-W48 Both show better performance .
5. image classification
HRNet-Wx-C:4 Analysis of three characteristic maps with different resolutions bottleneck layer , After doubling the number of channels , From the high-resolution image in turn strided
convolution Element addition with low resolution image , After 1*1 Convolution doubles the channel (1024->2048), After global average pooling, it is sent to the classifier .
experiment ：ImageNet Classification
And Resnet contrast
HRNets In and Resnets When the calculation amount of parameter is similar , Results and ResNets Quite , And slightly better than ResNets.
6. object detection
HRNetV2p: take HRNetV2 The stitched feature maps generate different levels of feature representation through average pooling operation of different scales , after 1*1 After convolution, the feature pyramid is formed
Technology