[Feature] It is hoped that RTMPose 3D will provide a lightweight model containing only 17 key points.

[mc373906408]

What is the feature?

Depth information is required for pose analysis, but I only need the key points of the limbs; outputting redundant key points will result in a waste of performance.

Any other context?

No response

[Homeworkbo]

Hi. I have also met this problem.
In my experiment, the redundant points might not cause too much waste of performance. The head part only takes a very small part in the total FLOPs (I think the relevant modules are "cls_x", "cls_y", and "cls_z" of the head, which only take less than 1% of the total FLOPs).
So I think it is OK to just mask the redundant keypoints of the output.
And, by the way, I think the best way is to fine-tune a new model with a different downstream head, as there exists a very serious bug in data augmentation. That is, when flipping the image, the depth coordinates of the left and right parts of the body will not exchange as the x and y axes do, which might cause inaccuracy in the depth axis.