[Feature] yoloxpose evaluating on bbox and not keypoints?

[cacticaxas]

What is the feature?

Hello,

i noticed that the validator in the yoloxpose and rtmo config were set to score_mode = bbox. Does this mean that the AP values receive through the tools/test.py script is only for the bounding boxes?

And if so, why is not set to "keypoints"?

Any other context?

No response

[XiongZhongxia]

I have the same question.

• Test: Data: coco pose val set Model config: rtmo-s_8xb32-600e_coco-640x640.py Model weight: rtmo-s_8xb32-600e_coco-640x640-8db55a59_20231211.pth

• Results: score_mode bbox: 0.677 (same as official results in https://github.com/open-mmlab/mmpose/tree/main/projects/rtmo) score_mode keypoint: 0.522 score_mode bbox_keypoint: 0.672

Is there any explaination for these results?

And most importantly, is the metric for the reported results of RTMO really aligned with the metric used in yolov8-pose (https://docs.ultralytics.com/tasks/pose/)?

I think it is very important to clarify this problem, since there may be a significant gap when using score_mode=bbox and score_mode=keypoint. I believe the result in the paper are reported with score_mode=bbox (https://openaccess.thecvf.com/content/CVPR2024/papers/Lu_RTMO_Towards_High-Performance_One-Stage_Real-Time_Multi-Person_Pose_Estimation_CVPR_2024_paper.pdf)

[Ben-Louis]

This parameter only influences the score of each person and determines which instances will be retained in multi-person scenarios. see https://github.com/open-mmlab/mmpose/blob/main/mmpose/evaluation/metrics/coco_metric.py#L453-L456