SAHI — Slicing Aided Hyper Inference — is an open-source computer vision library from OBSS that dramatically improves object detection on large images and small objects by running inference on overlapping slices of the image and merging the results. It has earned more than 5,500 stars on GitHub and is actively developed, with its latest updates in 2026.
Standard detectors downscale images to a fixed input size, which can erase small objects entirely — a serious problem for satellite imagery, aerial surveys, surveillance and industrial inspection, where the objects that matter may be a few dozen pixels across in a 4K or larger frame. SAHI slices the image into overlapping patches, runs your existing detector on each patch at full resolution, and intelligently merges the predictions back into the original coordinate space. It is framework-agnostic, working with YOLO-family models, MMDetection, Detectron2 and Hugging Face detectors, and extends beyond plain detection to instance segmentation, oriented bounding boxes and open-vocabulary detection. The library also ships an interactive UI and error-analysis plots for prediction inspection, plus COCO dataset utilities for slicing datasets at training time too. In published benchmarks on aerial datasets such as VisDrone and xView, adding SAHI to a standard detector yields large gains in small-object accuracy without retraining the underlying model.
SAHI is completely free and open source under the MIT license.
SAHI is for computer vision engineers working with high-resolution imagery — remote sensing, drones, surveillance, manufacturing inspection — where small objects matter and a vanilla detector misses them.