How watch-crop works

watch-crop finds the best requested object, builds a crop that preserves the requested aspect ratio, and returns an exact-size JPEG.

1 · DecodeRead the image bytes, correct EXIF orientation, and keep the original resolution.

2 · DetectRun YOLOv8x-OIV7 on the full frame at 640px. If no target is found, inspect overlapping tiles.

3 · FilterKeep only the selected classes, apply non-maximum suppression, and score confidence against object area.

4 · CropExpand the target by about 10%, then find the smallest requested-ratio rectangle that contains it.

5 · ResizeTranslate the crop inside the source bounds and resize it to the exact requested width and height.

What the response tells you

X-Target-Class, X-Detection-Confidence, X-Detection-Box, and X-Crop-Box describe the decision. The home page draws those two boxes over your original upload.

No machine learning is used after detection; crop geometry and resizing are deterministic.