yolo26n-catdog

A YOLO26n detector fine-tuned to find cats and dogs, exported to ONNX so it can run in the browser with onnxruntime-web. It is a learning exercise, not a production model.

Base model	yolo26n.pt (Ultralytics 8.4.171)
Classes	0: cat, 1: dog
Input	images — float32 [1, 3, 640, 640], RGB, 0–1, letterboxed (pad 114)
Output	output0 — [1, 6, 8400] = box cx, cy, w, h (pixels in the 640 frame) + 2 class scores. NMS is not included
Size	2.4M parameters · 5.3 GFLOPs · ONNX 9.3 MB (opset 17, simplified)
Training	100 epochs on 20 Roboflow-labelled images (train 14 · valid 4 · test 2), CPU
Result	best mAP50 0.9125 (epoch 76), last 0.874 — measured on only 4 validation images, so treat it as a smoke test

Limits

• Trained on 20 images. It misses many cats and dogs, especially in memes, crowded scenes, or small objects.
• Scores are low (often 0.2–0.4). Lower the confidence threshold to see more boxes.

Use

from ultralytics import YOLO
model = YOLO("yolo26n_catdog.onnx")
model.predict("image.jpg", imgsz=640, conf=0.25)

In the browser it is used by the model-card site AI Models Organize (YOLO26 card → «돌려보기» tab), which letterboxes the image, runs the session, and applies per-class NMS in JavaScript.

License

AGPL-3.0, inherited from Ultralytics YOLO.