how do you pick out coco_40k? about meshtransformer HOT 2 CLOSED

Great question! We filter the samples by the number of visible keypoints. If the sample have >=12 visible keypoints, we include the sample in our training, resulting in a total of ~40k samples. This is the same as GraphCMR. We follow their procedure.

We actually tried to use full MSCOCO, but we observed performance degradation. It is about ~3 PA-MPJPE worse. We think this is mainly because in our 2D-reprojection, we don't have penalty on the invisible keypoints. If the majority of the human body is out-of-the-view, it is free to hallucinate... and difficult to optimize 2D re-projection at least in our framework.

You may find out-of-the-view examples in the supplementary material of our Mesh Graphormer paper.

from meshtransformer.

Thanks for your quick and detailed reply! :)

from meshtransformer.