Im working on a project to detect keypoints on a massive dataset of computer generated football pitch images. I created a workflow to add annotations using coco keypoint json format. when the images are uploaded to the dataset tab, I can see that they have been annotated correctly - all annotations fit the key points as expected. - see below
I then proceded to training the keypoint detection model on these images which I had checked were correctly annotated with any non visible points deleted from the annotation skeleton.
When it came to training the model I noticed while sense checking some of the training images that they looked wrong. After looking through all the images I could see that the keypoints which are not visible in the image that were showing as deleted in the annotation phase are now being placed in the images for the trining dataset whcih completely messes up training. - see below
This has cost me a lot of training time and quite a large amount of wasted credits but also I’m not even sure how to troubleshoot the issue.
