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
4.31k stars 908 forks source link

higher result #174

Open wongkwan opened 4 years ago

wongkwan commented 4 years ago

Hello, I used your code to test the result of pose_hrnet_w32 on the coco val2017 dataset. Its input dimension is 256*192. Why I get the result 76.5 higher than the result 74.4 given by the original paper?

leoxiaobin commented 4 years ago

I think you used the ground truth boxes for evaluation. The results reported in paper are achieved by using the boxes from a person detector.

wongkwan commented 4 years ago

thank you so much. There is one more question I would like to ask you. I changed the network structure part of your HRNet to the network structure of CPN( Cascaded Pyramid Network ), I got 72.2 AP, but the CPN paper only had 69.4 AP, I would like to ask whether you have made optimization in the data preprocessing part, or what other reasons cause the CPN network structure to be better than the results in the paper?