Object detection (YOLO26) and binary segmentation (U-Net) pipelines for skeleton weed identification in drone imagery, built on fastai.

shared/              Shared data loading, image augmentations, helpers
detection/           YOLO26 object detection pipeline
segmentation/        U-Net binary segmentation pipeline
models/              Trained checkpoint outputs

uv sync

Set data_dir in each notebook's config cell to point to your local SKW dataset directory.

• 434 images (792x792), YOLO normalized bbox labels
• 2 classes: skw_0S, skw_1R (or single-class mode)
• Split: 389 train / 45 val / 2 test

YOLO26 (ultralytics) wrapped for fastai with E2E loss, EMA, and mosaic augmentation.

# Interactive notebook
cd detection && uv run jupyter notebook train.ipynb

# Script training
uv run python detection/train.py

# Hyperparameter search
uv run python detection/optuna_search.py

smp U-Net with ConvNeXt encoder. YOLO bbox labels are rasterized into binary masks for pixel-wise training. ForegroundIoU is directly comparable to detection's RasterIoU.

cd segmentation && uv run jupyter notebook train_seg.ipynb

uv run ruff format .
uv run ruff check . --fix