Hi, I have a dataset that contains pictures, and some of them have yellow balls in them. Basically I want to find the center point of each of the balls. There could be anywhere from 0-8 balls in the dataset, although when I use it, there is no actual cap on the number of balls in the frame. I was thinking I would need a regressive model for this, but also multiclassification, but I don’t really know if I can do multiclassification when some elements of the dataset don’t even have balls in them.
Any help would be great!
Sounds like object detection might be what you’re after. It’s a topic that was covered in previous courses - https://www.youtube.com/watch?v=0frKXR-2PBY&t=3840s.
Alternatively you might want to look up facial keypoints detection for an example of regressing on multiple points - https://www.kaggle.com/nitron/facial-keypoints-fastai-image-regression/code
Is there any pattern in the relationship between the balls? Not sure how much the facial keypoints stuff works as a result of learning about where things on a face should be but I guess you can always just try it and see what results you get.
No, there really isn’t any pattern to the balls, thank you for the advice though. I think you are right about the object detection
The facial keypoints is definitely something to consider. Otherwise mixing some sort of object detection model on top of a keypoint model to find your points would be a solid option too (consider HRNet as one example)