sahi

Why we included this project

SAHI tackles a common pain point: object detectors that fail on small objects spread across very large imagery. Instead of changing the model, it slices the image into overlapping tiles, runs inference on each piece, and stitches the predictions back together. That makes small objects appear larger to the detector and lets you process satellite and aerial imagery, as well as high-resolution street scenes, that would otherwise exceed GPU memory. It works with popular PyTorch backends like Ultralytics YOLO and MMDetection, and it also supports Hugging Face object detectors, so you can drop it into an existing pipeline without switching frameworks. Beyond inference, it supports fine-tuning on tiled data, and it includes utilities to slice COCO datasets and compute classwise metrics, plus error-analysis plots. Teams doing remote sensing or any large-image detection work will find it a practical, battle-tested companion.