I'm lanning to create a real-time wall damages detector [scratches, Cracks] using YOLOv5 and my custom dataset of images (125).
For now I'm just trying to do a proof of concept. Wanted to have my steps planned in advance.
If your dataset is small (like in you case), transfer learning will almost always give better results when compared to training from scratch. As for your second question, yes. The more data you get, the better your model will be able to learn and perform. Considering that it's a relatively different task than which Yolo-V5 was originally trained on, try to get as many images as you can