Create segmentation masks with a U-net

[andandandand]

Training the model

We will use the subset of approximately 100 images that we have manually tagged (with help of Segment Anything 2) inside CVAT to train, validate, and test the U-net model.

Use 50 images for training, 25 for validation, 25 for testing.

Share your training notebooks on the Github repository and help each other complete the task.

Repository: https://github.com/JaguarIdentification/SegmentationSprint

Claude's guide to adapt the U-net from FastAI with with CVAT (COCO 1.0 format)

https://claude.ai/public/artifacts/532e6f6e-e29e-4e43-9e50-08643bc7d0c7

U-Net Tutorial Resources

Basic U-Net Implementation

• Tutorial Notebook: [U-net Tutorial](https://colab.research.google.com/github/andandandand/practical-computer-vision/blob/main/notebooks/U-net_Tutorial.ipynb)

FastAI Implementation

• Example Notebook: [FastAI2 Dynamic U-Net](https://colab.research.google.com/drive/1iEU10Te1Kgk_t_w7NQ2H55fsyEtXkCz2?usp=sharing)

Performance Note: The FastAI2 example achieves better performance with less training due to several optimizations:

• Pre-trained ResNet backbone
• Differential learning rates
• One-cycle training policy

Documentation: Using FastAI requires familiarity with their DataBlock and DataLoader syntax. See the [DataBlock tutorial](https://docs.fast.ai/tutorial.datablock.html) for reference.