What I built
- Training pipeline from raw drone imagery to a fine-tuned segmentation model.
- Data augmentation aware of aerial-specific issues (shadows, herd overlap, motion blur).
- Inference notebook with per-frame metrics so the rancher can audit the count.
What I learned
- In agricultural CV, the data pipeline is harder than the model. Models are mostly solved.
- Domain experts spot mistakes the eval metric never will. Always show predictions to the rancher.
- PyTorch is the right default. The ergonomics of writing custom loss functions justify the verbosity.
Stack
- Python
- PyTorch
- OpenCV
- NumPy
- Albumentations
- Jupyter