Generating segmentation masks Automaticly via a transfer learned CNN

Are there any FastAI resources on using maybe a CNN to automatically make segmentation masks? I only know how to use the FastAI libraries for basic things and have never built any Deep learning architectures reproduced a paper.

Sure, check out the end of the vision tutorial:

I appreciate the link. However, I think I need to clarify. I’m trying to find a way to generate novel image segmentation masks for categories that do not have labeled data associated with them via some CNN that has learned the categories via transfer learning. The closest thing I’ve found is Fully Convolutional Networks for Semantic Segmentation . But, I am trying to understand the code examples and read the paper.

Oh, ok, this is more advanced. Maybe Facebook’s DINO is what you are looking for? It does mostly use Vision Transformers not CNNs. However it generates very accurate attention maps, similar to a binary segmentation mask.

They have published all Code on GitHub:

Thanks a lot! I found a FastAI version. I’m hoping I can use this to generate masks and then train an instance segmentation algorithm.

1 Like