PinataFarms / DAD-3DHeads

need retraining?

It is sensitive to the way you crop the head, indeed. If you want to have the model work on your custom crops, you would likely need to retrain the model for your particular use case. The training code is available in the repo.

The training dataset contains bounding boxes for heads. You can see here

DAD-3DHeads/model_training/data/flame_dataset.py

Lines 97 to 98 in bfb7371

    
           offset = tuple(0.1 * np.random.uniform(size=4) + 0.05) 
        
           x, y, w, h = ensure_bbox_boundaries(extend_bbox(np.array(bbox), offset), img.shape[:2])

how we pre-process those bounding boxes during the training stage.

Hope this helps!

	offset = tuple(0.1 * np.random.uniform(size=4) + 0.05)
	x, y, w, h = ensure_bbox_boundaries(extend_bbox(np.array(bbox), offset), img.shape[:2])

it is not so good, it is good only for the 1.jpeg image.