Model: huggingface.co/Isa0/cat-detection Dataset: CAT Dataset – Kaggle

Detects 9 facial landmarks on cats (eyes, ears, nose, mouth) and draws a bounding box around them.

Two variants, both mobile-friendly (well under 100 ms per inference on a phone):

Both models take RGB input in [0, 1] (ImageNet normalization is baked into the exported graph) and output 18 values: 9 (x, y) landmark coordinates normalized to [0, 1] in the letterboxed square input space.

train.py implements:

• Aspect-ratio-preserving preprocessing — images are padded to square on both sides of the short edge (letterbox) instead of being stretched, so the cat's facial anatomy is never distorted. Inference (main.py) uses the same letterboxing and maps predictions back to original image coordinates.
• Augmentations — random horizontal and vertical flips (with correct left/right relabeling of the symmetric eye/ear landmarks) and random rotation in all directions (−180°…180°, canvas expanded so nothing is cut off).
• Random crop samples — each source image also contributes a randomly cropped version whose crop edges always stay outside the annotated landmark bounding box plus a margin, so facial points are never lost.

uv sync --group train
python train.py --data-dirs data/CAT_00 data/CAT_01 ... --variant both --epochs 30

--variant small|large|both selects which model(s) to train. Best-validation weights are checkpointed to cat_landmark_{variant}.pth and exported to cat_landmark_{variant}.onnx.

python main.py test.jpg --variant small          # or --variant large
python main.py photo.jpg --model path/to/model.onnx --output annotated.jpg

Outputs landmark coordinates to stdout and saves an annotated image.

• onnxruntime
• opencv-python
• training additionally needs torch, torchvision, onnx (see the train dependency group)

Install with uv:

uv sync