Skip to content

Quick Start

SAHI (Slicing Aided Hyper Inference) detects small objects in large images by slicing them into overlapping tiles, running your detector on each tile, and merging the results. It works with any detection model no retraining needed.

sliced inference

Installation

PyPI - Version Conda Version PyPI - Python Version

pip install sahi

For object detection you also need a framework. The most common choice is Ultralytics:

pip install ultralytics
Other install methods

Conda:

Conda Downloads Conda Platforms

conda install -c conda-forge sahi

Note

If you are installing in a CUDA environment, it is best practice to install ultralytics, pytorch, and pytorch-cuda in the same command: 

conda install -c pytorch -c nvidia -c conda-forge pytorch torchvision pytorch-cuda=11.8 ultralytics

From source: 

pip install git+https://github.com/obss/sahi.git@main

Development (editable): 

git clone https://github.com/obss/sahi
cd sahi
pip install -e .

See the pyproject.toml for the full list of dependencies.

Sliced Prediction with Python

from sahi import AutoDetectionModel
from sahi.predict import get_sliced_prediction

# Load a model (works with any supported framework)
detection_model = AutoDetectionModel.from_pretrained(
    model_type="ultralytics",
    model_path="yolo26n.pt",
    confidence_threshold=0.25,
    device="cuda:0",  # or "cpu"
)

# Run sliced prediction
result = get_sliced_prediction(
    "path/to/your/image.jpg",
    detection_model,
    slice_height=512,
    slice_width=512,
    overlap_height_ratio=0.2,
    overlap_width_ratio=0.2,
)

# Export visualizations
result.export_visuals(export_dir="demo_data/")

# Access individual predictions
for pred in result.object_prediction_list:
    print(pred.category.name, pred.score.value, pred.bbox.to_xyxy())

Prediction with the CLI

Run sliced inference without writing Python code:

sahi predict \
  --model_path yolo26n.pt \
  --model_type ultralytics \
  --source /path/to/images/ \
  --slice_height 512 \
  --slice_width 512

Results are saved to runs/predict/exp by default.

Choosing a Postprocessing Backend

After slicing, SAHI merges overlapping predictions with NMS or NMM. The best available backend is selected automatically:

Backend When selected Install
torchvision CUDA GPU + torchvision available pip install torch torchvision
numba numba installed, no GPU pip install numba
numpy Always available (fallback) --

Override the choice manually:

from sahi.postprocess.backends import set_postprocess_backend

set_postprocess_backend("numpy")       # always available
set_postprocess_backend("numba")       # JIT-compiled
set_postprocess_backend("torchvision") # GPU-accelerated
set_postprocess_backend("auto")        # restore auto-detection

Next Steps