I’m trying to figure out the appropriate way to train a MaskRCNN model that has key points.
The basis for my approach is here:
You can see that in that approach where we fine tune the model, we add a fastrcnn and a MaskRCNN predictor to the ROI Head. In this case we are fine tuning both masks and bounding box.
What I have is a complete different dataset that contains key points - for pose estimation. The reason is simple, Coco provide masks, bounding box, labels and key points for humans, but all other classes only have masks, bounding box and label.
So, how can I add a keypoint predictor to the ROI Head and only train that? Does anyone have experience with this?
Appreciate and advice.