Closed xins981 closed 5 months ago
The viewpoints generated from the object coordinates cannot be strictly matched to the ones from the camera coordinates one by one. So the grasp labels at one viewpoint generated from the camera coordinates (in training) are estimated by searching the nearest viewpoints from the object coordinates, where the original labels are in.
https://github.com/graspnet/graspnet-baseline/blob/fb2f089a9515d2a5e433b28b53827c1c914437f2/utils/label_generation.py#L60 In my opinion, grasp_viewstrans are the viewpoints in camera frame, rather graspviews are ones in object or world frame. What is the goal of knn operation?