obss/sahi

What is it?

Key features

• 01Sliced inference that runs detection on image tiles and merges results back to full-image coordinates.
• 02Unified API and CLI across Ultralytics, MMDetection, Hugging Face, TorchVision, Detectron2, YOLOv5 and RF-DETR models.
• 03Video and image prediction from the command line.
• 04COCO utilities for slicing, subsampling, filtering, merging, splitting and converting to YOLO format.
• 05Classwise COCO AP and AR evaluation and error analysis plot export.
• 06Interactive result inspection through the FiftyOne app.

Use cases

• Detecting small objects such as vehicles or people in high-resolution aerial or satellite imagery.
• Running inference on large surveillance or industrial inspection images without downsampling.
• Converting and slicing COCO datasets before training a YOLO model.
• Analyzing where a detector fails by generating error analysis plots.

Quick start

1. 1

   Install SAHI from PyPI

   pip install sahi
2. 2

   Install PyTorch and torchvision for your hardware (see Detailed Installation in the README)

   pip install torch==2.13.0 torchvision==0.28.0 --index-url https://download.pytorch.org/whl/cu130
3. 3

   Install a detection framework such as ultralytics

   pip install ultralytics>=8.3.161

Things to know

• !Full use requires installing a separate detection framework and PyTorch, which adds setup effort.
• !Sliced inference runs the detector over many tiles, so it is slower than standard prediction.

In collections

Topics

FAQ

What is sahi?

A Python library for sliced inference that runs object detection models over tiles of large images to find small objects. SAHI (Slicing Aided Hyper Inference) addresses a common problem in object detection: small objects become hard to detect when a large image is resized to a model's input size. It splits images into overlapping slices, runs detection on each slice, and merges the results back into full-image coordinates.

Is sahi free to use? What license does it have?

sahi is open source under the MIT license. Check the license terms before using it in commercial products.

Is sahi still maintained?

Actively maintained. The last commit to sahi was on Nov 1, 2026.

What can I use sahi for?

Detecting small objects such as vehicles or people in high-resolution aerial or satellite imagery. Running inference on large surveillance or industrial inspection images without downsampling. Converting and slicing COCO datasets before training a YOLO model. Analyzing where a detector fails by generating error analysis plots.

How do I install sahi?

Install SAHI from PyPI: pip install sahi Install PyTorch and torchvision for your hardware (see Detailed Installation in the README): pip install torch==2.13.0 torchvision==0.28.0 --index-url https://download.pytorch.org/whl/cu130 Install a detection framework such as ultralytics: pip install ultralytics>=8.3.161

Where can I get sahi?

The source code and installation instructions are on GitHub: https://github.com/obss/sahi