open-mmlab
How to set positive target_weigth and zero target for hidden keypoints
Hello,
I'm training a keypoints detector for a use-case in which I have 8 keypoints and for each sample I know in advance that 4 of the 8 keypoints are occluded. I set up the trainig dataset and trained a network, however during inference the network always predicts all 8 keypoints, instead of being able to recognize between the visible ones and the occluded ones.
Tracking down the issue, I realized that in TopDownGenerateTarget._msra_generate_target() the target_weight is derived directly from the keypoint visibility, so the network receives weight updates only if a keypoint is present in the image.
What I want to enable is for the network to have negative feedback too, by setting the target_weight to 1 and having an all-zero target heatmap.
Is there a way to enable this behaviour? Or should I implement my own variation of TopDownGenerateTarget._msra_generate_target()?
Is there a way to enable this behaviour? Currently, no.
_Or should I implement my own variation of TopDownGenerateTarget.msra_generate_target()? Yes. I think so.
