The pose estimation on ground truth bbox is worse than on detected bbox, at Test-Set.

[ArchNew]

I use both HRNet model trained on COCO and trained on CrowdPose to evaluate on Test-Set. Both models have the problem that their pose estimation results using ground truth bounding box are worse than the results using detected bounding box. Not worse a little, but a lot. Especially for the model trained on COCO, up to 20% difference. Evaluated with ground truth bbox, I have 58.6% mAP, but with detected bbox, I can get 76.6% mAP.

[ArchNew]

I found the issue. HRNet's strict filtering is detrimental to the ground truth bbox.