If I use the video 'spider.mp4' (480 * 640 Px) the skeleton is predicted nicely but with a shift to the actual spider (c.f. video 'spider_posture.mp4')
If I use the video 'spider_copped.mp4' (384 * 384 Px) instead where height=width the skeleton fits the actual spider (c.f. video 'spider_cropped_posture.mp4'; skeleton prediction messes up when only parts of spider are on the video - but this is not my point here.)
If I use the video 'spider.mp4' (480 * 640 Px) the skeleton is predicted nicely but with a shift to the actual spider (c.f. video 'spider_posture.mp4')
If I use the video 'spider_copped.mp4' (384 * 384 Px) instead where height=width the skeleton fits the actual spider (c.f. video 'spider_cropped_posture.mp4'; skeleton prediction messes up when only parts of spider are on the video - but this is not my point here.)
You find the data here: https://github.com/urs-waldmann/spider-data-for-DPK You find the modified notebooks for 'copped' videos here: https://github.com/urs-waldmann/DeepPoseKit/tree/spider You find the modified notebooks for 'original' videos here: https://github.com/urs-waldmann/DeepPoseKit/tree/issue_img_res