Exo-View Hand Key Point Annotations


I have some questions regarding 2d hand key point annotations. The dataset has 2D hand key point annotations on the exo-views, even when the hands are completely occluded.

Because there is no separate section dedicated to this part in the paper—https://arxiv.org/pdf/2311.18259, this part is a bit unclear to me, so could you please enlighten me on this?

Based on the appendix 13.D.2. 3D human hand pose from ego-video, I am guessing that, in the dataset, the 2D hand key points on the exo-views with the label “automatic” are purely estimated by off-the-shelf body pose estimation models. Thus, the annotations are there even when the hands are occluded. would it be correct?

For the key points with the “manual” label, we are certain that they are at least visible and the annotators were able to annotate them, right?

Also, some annotations are really off from other given key points. Would there be any reason or suggested way to filter them out?


Thank you so much for the clarification in advance!