SAHI-style tiling for YOLO, minus the friction

What it does

YOLO-Patch-Based-Inference is a Python library that brings SAHI-style tiling to the Ultralytics ecosystem. It slices high-resolution images into overlapping crops, runs detection or instance segmentation on each patch, then merges the results with custom NMS logic. It supports a broad roster of Ultralytics models—from YOLOv8 through YOLO12, plus FastSAM and RTDETR—and wraps the whole flow in a two-class API.

The interesting bit

Instead of treating NMS as a dumb confidence filter, the library offers an “intelligent sorter” that bins detections by area and rounded confidence, so a large mediocre box can outrank a tiny high-confidence artifact. The README also advertises “sleek customization” for visualizing both standard and patch-based results.

Key highlights

• Broad model support: YOLOv8–YOLO12 (detection and segmentation), FastSAM, and RTDETR, using pre-trained or custom weights
• Two-step design: MakeCropsDetectThem handles tiling and inference; CombineDetections merges overlapping predictions with configurable IOU or IOS matching
• Batch inference mode for processing crops in parallel on GPU, trading VRAM for speed
• Optional memory optimization for segmentation, though the README warns this sacrifices some accuracy
• Built-in visualization helpers for both standard and patch-based results

Caveats

• Resizing results back to the original image size is flagged as a slow operation
• The segmentation memory-optimization mode explicitly trades accuracy for lower RAM usage
• Tightly coupled to the Ultralytics model format; not a general-purpose tiling framework

Verdict

Worth a look if you are already in the Ultralytics YOLO orbit and need to detect small objects in large images without rewriting SAHI glue code. If you are not using YOLO-family models, this library will not help.