search

leoxiaobin / deep-high-resolution-net.pytorch

The project is an official implementation of our CVPR2019 paper "Deep High-Resolution Representation Learning for Human Pose Estimation"
https://jingdongwang2017.github.io/Projects/HRNet/PoseEstimation.html
MIT License
4.32k stars 913 forks source link

Why heatmap size is also downsample into 4 times smaller #31

Open KeepGoing2019HaHa opened 5 years ago

KeepGoing2019HaHa commented 5 years ago

Hi, may I ask why heatmap size is also 4x smaller than input image? I mean if you are keeping high resolution feature maps all the time, why do not just generate heatmaps with original input size for training and inference? Is there something that I am missing?

YinRui1991 commented 5 years ago

I think it's probably empirical. I find some other codes also use 4x smaller heatmap. I think it's OK to use other ratio, but the model should be trained with new output size.

KeepGoing2019HaHa commented 5 years ago

@YinRui1991 yes, almost all pose estimators use 4x downsample. But this paper aims to maintain high resolution feature maps, why not also generate high resolution heatmaps?

sunke123 commented 5 years ago

@KeepGoing2019HaHa In this paper, we focus on maintaining the high-resolution feature maps, not increasing the resolution. So, we maintain the high-resolution feature maps as used in most works, 4x feature maps. You can try to use higher resolution feature maps, such as 2x or full-resolution. I think that is also helpful. Thanks for your attention.

HuAndrew commented 5 years ago

I guess their work is different from other network, like hourglass, simple baseline. In 2018 ECCV 2018 simple baseline papers, feature maps is smaller than 4x feature maps. Using 4x feature maps fusion is well to balance parameters and performance.

aliYaHf commented 5 years ago

@sunke123 hi ,Can you provide training logs on the MPII dataset? I don't know why the AR value of my training is very low.

ghost commented 4 years ago

@sunke123 Hello,is the groundtruth heatmap's resolution also 4 times smaller than input image's resolution? If what I think is true,how can I modify the groundtruth heatmap's resolution? Looking forward to your reply.