SAHI

Small objects get lost when a high resolution image is squeezed down to a detector's input size, and SAHI attacks that problem by slicing the image into overlapping tiles, running inference on each slice, and merging the predictions back into full-image coordinates. The approach, called slicing aided hyper inference, works as a wrapper around existing models rather than a new architecture, with adapters for Ultralytics YOLOv8 through YOLO26, MMDetection, Detectron2, TorchVision, Hugging Face Transformers, RT-DETR, Roboflow RF-DETR, YOLO-World, and YOLOE. The package installs with pip install sahi, adds a CLI for prediction plus COCO dataset utilities such as slicing, subsetting, and error analysis, and integrates with FiftyOne for visual inspection. It runs wherever the underlying detector runs, on CPU or GPU, and is released under the MIT license. The method was published at IEEE ICIP 2022 and is widely used for aerial and satellite imagery, surveillance footage, and other domains where targets occupy a handful of pixels; the repository has passed 5,000 GitHub stars.