sahi-inference

Tiled inference applies SAHI's get_sliced_prediction to split large drone orthomosaics into overlapping tiles, run object detection on each tile, and merge results back to full-image coordinates to catch small targets.

SAHI-based tiled inference detects wildlife in drone orthomosaics by slicing images into overlapping tiles, running detectors on each, and merging predictions using NMS or NMM to resolve duplicates and dense clusters.

Yes, SAHI's AutoDetectionModel backend supports YOLO, Detectron2, and MMDet for sliced inference on satellite tiles. You use get_sliced_prediction to process large images and merge overlapping tile predictions.

You need tiled inference when processing very large images like drone orthomosaics or satellite tiles where single-pass object detection misses small targets. SAHI splits images into overlapping tiles to ensure full target visibility.

NMM merging resolves dense clusters of objects in drone survey predictions, whereas NMS simply suppresses overlapping duplicates. NMM is better suited for crowded wildlife scenes where multiple targets overlap across tiles.

Yes, MegaDetector can be wrapped with SAHI for wildlife monitoring from drone surveys. The tiled inference pipeline supports MegaDetector wrapping to produce GIS-ready outputs for census, monitoring, and biodiversity surveys.