Fish segmentation

A fish segmentation pet-project. U-Net and pretrained DINOv2 architectures were tested.

Tech stack:

• DVC for experiment tracking
• pytorch
• pytorch lightning
• Albumentations
• Gradio

Dataset

SUIM Dataset.

For this project only fish masks were sampled.

964 Train-val datapoints were randomly sampled in 80% to 20% ratio. Test dataset containes 66 annotated images.

Additionaly 10% images without fish were added to train, val and test splits.

Augmentations

Train-val set was augmentated using Albumentations library:

• ShiftScaleRotate
• RGBShift
• RandomBrightnessContrast
• HorizontalFlip

Results

The best model was U-Net with IOU 0.598 on test dataset.

Demo

Demo can be accessed running

python scr\demo.py