Rejection rule in correspondence key-points estimation

[zohaibmohammad]

Hi,

Thank you for making your work public. I am stuck in a problem, can you please help me? How could I reject the extra (not required) key-points especially in test time? For example, during training, I can use "-1" to exclude the extra key-points but in test scenario all (21) key-points are estimated by the model. So, how could I ignore the key-points that are not required by assuming that I don't have any info about ground truth.

Best regards, mz

[qq456cvb]

We skip those keypoints that do not exist, which is a common practice, e.g. in Deep Functional Dictionaries (https://github.com/mhsung/deep-functional-dictionaries/blob/master/network/evaluate_keypoints.py#L35). If you don't want any info of ground-truth to be leaked, you may estimate an additional binary mask predicting the existence of keypoint. Then any inconsistent existence prediction results in an infinity distance to the ground-truth. This may require a small modification on the existing code.