sahi-inference

What problem does it solve?

SAHI is a tiled inference framework that enables robust object detection on very large images by splitting them into overlapping tiles, running detectors on each tile, and merging results to full-image coordinates. This approach is essential for wildlife monitoring from drone orthomosaics and satellite imagery where small targets can be missed in a single pass.

Core Features & Use Cases

• Slice-based inference: operate on oversized images by tiling with controlled overlap to ensure full object visibility.
• AutoDetectionModel integration and get_sliced_prediction usage: plug-in YOLO, Detectron2, MMDet, or other backends with SAHI for scalable workflows.
• Postprocessing options: NMS and NMM merging strategies to resolve duplicates and dense clusters; supports drone orthomosaic pipelines.
• Integration with drone/image pipelines: supports MegaDetector wrapping and GIS-ready outputs for census, monitoring, and biodiversity surveys.

Quick Start

Run a sliced inference on a large orthomosaic using SAHI with a YOLO model.