SAHI is a lightweight vision library for performing large-scale object detection and instance segmentation on high-resolution images through sliced inference.
The tool addresses the challenge of detecting objects in large images where standard detectors struggle with small objects or memory constraints. SAHI solves this by dividing images into overlapping slices, running inference on each slice independently, and then merging the results intelligently. This sliced inference approach allows models to maintain their effectiveness on small objects while processing images that would otherwise exceed memory limits or degrade detection quality.
The project is framework-agnostic, supporting multiple detection backends, and includes an interactive user interface for visualization and error analysis. It handles oriented object detection, open-vocabulary detection, and remote sensing applications. Developers working with satellite imagery, aerial photography, or any scenario involving large images with small objects should consider SAHI. Teams needing to analyze detection failures and understand model behavior will benefit from the built-in error analysis plots and interactive tools.
The project maintains active continuous integration with security monitoring. Development follows a structured approach to code quality and testing. The tool has been formally published in academic literature, indicating sustained validation of its approach beyond typical open-source development cycles.
