Hello, I am trying to apply transfer learning to a YOLO26n-pose model for it to learn a new class with no keypoints, basically treating it like object detection. I cannot use two models for my architecture and must run on the pose. When I try to train my model, I freeze all layers except the head (I apply freeze 22). My model begins to learn the new class, but the person class is either not showing up or shows up in stiff positions (positions the people in the data were in) and are not fluid like how they are from the out of the box model that I trained on. I labeled people in my dataset. I was wondering how to resolve this issue.
There’s no practical way to avoid performance degradation when you fine-tune on a different dataset, other than also adding images from person class in your dataset. However, increasing class number will cause the head to reset so the head will train from scratch which is why you will see significant reduction in accuracy.
If I add data from the person COCO dataset that the model was trained on, roughly 5k images, will that resolve this issue, or will it drown out the other class during training that has 60 images.
It will help with person class. But you should have more images for your custom class. 60 images isn’t enough for training a good model.