Hi! I’m working on Object Detection training on my labels from COCO checkpoint aka Transfer Learning
I’m using Roboflow 3.0 (instant, accurate) models and looking at YOLO (fast, accurate) 10/11/12.
My question: how many labeled images i need for good convergence per each class?
Deeper question: what type of transfer learning are used for these models == how many DNN parameters we are changing. Knowing that i can predict if 30 or 100 or 300 images per class needed to fine-tune X parameters.
There isn’t an exact number of images because it is highly dependent on the problem you are solving. We typically recommend training your first model when each class has at least 50 images with representation.
Then train a model with just Resize and Auto-Orient, and see how performance is, using our Model Evaluation tools.
After that you can see what classes the model is struggling with, and introduce progressive preprocessing and augmentation steps to see how it impacts performance!