What are some useful rules of thumb for picking the number of augmenters per training image?

I realize this is a hyperparameter I can vary and test: I'm just trying to get a sense for reasonable ranges/choices based on people's experience with different network architecture and other particulars.

My particulars are that I have a relatively simple object detection task, with about 1100 annotated images, a single category. I am doing the transfer learning on a faster-rcnn with resnet-101 backbone.

I do an 80/20 train/test split, and even when I augment just 5x for each training image [h/v flips, small rotations, brightness changes, gaussian blur, motion blur, scale], the network is learning fairly well given all my other parameters.

For my final step I am planning to just add more augmented images per training image. I had always though 100 was a good number but I now realizes this is sort of arbitrary, and I picked it up just from some random example I used once from the internet (!). Even with 10 augmenters per image, I would have over 8k training images for the category, which should be enough, no?

Note a few people have asked about how many images there should be per class for training, but they don't take augmentation into account.

Related questions


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.