HowieMa
commented
4 years ago Hi, thanks for your asking. Yes, you are correct. We will loss some limb structure in this setting, thus I think it may affect the pose module, but this is the only thing can do. That's why in my paper, I only use the G1 method on OneHand10K, instead of G1+6, as we know that G6 may be broken for many fingers. For example, if the middle point of one finger is invisible, the lime structure of this finger will be broken into two parts.
For the OneHand10K dataset, I just want to demonstrate my algorithm is also efficient on different dataset, as the this dataset is the few 2D hand pose datasets I can use. For a better understanding of it, it's better to ask the owner of it.
Besides, I think the invisible keypoint issue is a system error, which is introduced by the OneHand10K dataset. If this dataset itself does not care about invisible keypoints, I think there is no need to spend much time to fix this issue on it.
For your last question, "how can we get ....", it's really difficult to say, as this is the goal for all Computer Vision algorithm, a better and stabler method. It's better to refer to the latest CVPR/ICCV/ECCV papers to find better solution :).
Again, thanks for your kindly asking. Hops this can solve your concern.
It says Gth of limb representatios to the missing keypoints are set to zero maps, the limb structure in the missing keypoints will be missing too. Will it affect the subsequent pose mudule regression? How can we get the better and stabler hand keypoints you prefer?