YoungXIAO13
commented
4 years ago Hi,
- I think training from scratch on 50 pictures has a great probability to overfit on training data, thus a pre-training on a larger datasets (MS-COCO) and transfer-learning techniques could be considered.
- Besides, you could adopt some data augmentation techniques such like cropping objects and paste them to compose novel training samples, which can enrich the training data at some extent.
- Moreover, if you can have access to additional unlabelled data, then some other works like weakly-supervised object detection or semi-supervised object detection could also be considered.
zengwb-lx
I'm going to train a object detection model， But my training set only has 50 pictures，Can you give me some good advice，