This PR adds the possibility of training CTD models with generative sampling without providing conditional keypoints in the json annotation files (e.g. on COCO). If no 'cond_kpts' key is provided in the training annotations, it will:
compute bbox overlaps between all individuals in a given image
treats overlapping individuals as near pose instances
uses GT poses to add noise sampled from the generative sampler (including the near individuals for swapping errors)
in order to calculate conditional keypoints during training.
This PR adds the possibility of training CTD models with generative sampling without providing conditional keypoints in the json annotation files (e.g. on COCO). If no 'cond_kpts' key is provided in the training annotations, it will: