search

Res2Net / Res2Net-Pose-Estimation

Res2Net for Pose Estimation using Simple Baselines as the baseline
https://mmcheng.net/res2net/
36 stars 5 forks source link

测试结果有问题 #1

Closed lxtGH closed 5 years ago

lxtGH commented 5 years ago

这个用了Ground Truth BBOX，结果虚高，希望纠正。

gasvn commented 5 years ago

We follow the simple baseline to use their evaluation method. We intend to prove the effectiveness of Res2Net. You are welcomed to provide other avilable evaluation methods.

lxtGH commented 5 years ago

See this issue.https://github.com/microsoft/human-pose-estimation.pytorch/issues/55 You should set the USE_GT_BBOX opinion False in the config file https://github.com/gasvn/Res2Net-Pose-Estimation/blob/master/experiments/coco/res2net/res2net101_4s_26w_256x192_d256x3_adam_lr1e-3.yaml

gasvn commented 5 years ago

Thanks, it seems the author's config is not the config used to reproduce the reported results. Still, according to the link you provided, the Res2Net based method outperform the baseline when using gt box. pose_resnet_50 | 0.724 | 0.915 | 0.804 | 0.697 | 0.765 pose_res2net_50 | 0.737 | 0.925 | 0.814 | 0.708 | 0.782 we will update the performance based on both gt box and the person detector's results ASAP.

gasvn commented 5 years ago

We have updated the results based on both gt box and the person detector's results. Thanks again for your notification.