YOLO-Patch-Based-Inference

Why we included this project

Small objects are the classic failure case for object detectors: by the time a tiny target is downscaled to fit the network input, it has often shrunk to a few pixels and gets missed entirely. This library handles that the way SAHI does, by slicing the image into overlapping patches and running inference on each one before merging the results with non-maximum suppression. It wraps Ultralytics models, so you can point it at YOLOv8, YOLO11, YOLOv9, RTDETR, or FastSAM checkpoints, including custom-trained ones, and it covers both detection and instance segmentation. If your work involves aerial imagery, satellite data, or high-resolution industrial photos where targets are small relative to the frame, the pip-installable package and the included Colab notebooks are a practical starting point. The library also has its own visualization helpers for both direct and patch-based runs, which is handy when you want consistent output styling across experiments.