A fish segmentation pet-project. U-Net and pretrained DINOv2 architectures were tested.
Tech stack:
- DVC for experiment tracking
- pytorch
- pytorch lightning
- Albumentations
- Gradio
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.
Train-val set was augmentated using Albumentations library:
- ShiftScaleRotate
- RGBShift
- RandomBrightnessContrast
- HorizontalFlip
The best model was U-Net with IOU 0.598 on test dataset.
Demo can be accessed running
python scr\demo.py
