SAHI (Slicing Aided Hyper Inference)

SAHI is an open-source vision library designed for large-scale object detection and instance segmentation, specifically optimized for detecting small objects. It uses a slicing-aided methodology that allows users to improve detection accuracy without requiring additional fine-tuning of the underlying model.

SAHI is compatible with a wide range of popular object detectors, including:

• Ultralytics (YOLOv8, YOLO11, YOLO26)
• HuggingFace Transformers (detection and segmentation)
• RT-DETR, TorchVision, MMDetection, Detectron2
• YOLOv5, YOLOE, YOLO-World, and Roboflow RF-DETR

SAHI implements sliced inference to detect small objects in large images (e.g., 4K drone or satellite photos) that would otherwise be lost during standard model resizing. The workflow follows three steps:

1. Slice: The image is divided into a grid of overlapping tiles. This ensures objects at tile boundaries are captured in full in at least one patch.
2. Detect: The detector runs on every tile. By default, SAHI also runs the detector on the full image (perform_standard_pred=True) to catch large objects that might be split across tiles.
3. Merge: Tile-level predictions are mapped back to full-image coordinates. Overlapping detections are merged or suppressed using post-processing algorithms.

SAHI provides specialized Google Colab notebooks for performing sliced inference with several popular computer vision frameworks. You can find ready-to-run demos for:

• Ultralytics (YOLO series): YOLO26, YOLO11, YOLO11-OBB, YOLOv5, YOLOX, and YOLOE.
• Detection Frameworks: RT-DETR, RT-DETR v2, GroundingDINO, MMDetection, Detectron2, and TorchVision.
• Other Integrations: Roboflow / RF-DETR and HuggingFace models.

For YOLOX, a live demo is also available on HuggingFace Spaces.

Use Facebook's Detectron2 models for detection and instance segmentation using model_type="detectron2". Requires config_path.

pip install detectron2

from sahi import AutoDetectionModel
from sahi.predict import get_sliced_prediction

detection_model = AutoDetectionModel.from_pretrained(
    model_type="detectron2",
    model_path="path/to/model_final.pth",
    config_path="path/to/config.yaml",
    confidence_threshold=0.25,
    device="cuda:0",
)

result = get_sliced_prediction(
    "image.jpg",
    detection_model,
    slice_height=512,
    slice_width=512,
)

Load classic YOLOv5 models using the yolov5 pip package and model_type="yolov5".

pip install yolov5

from sahi import AutoDetectionModel
from sahi.predict import get_sliced_prediction

detection_model = AutoDetectionModel.from_pretrained(
    model_type="yolov5",
    model_path="yolov5s.pt",
    confidence_threshold=0.25,
    device="cuda:0",
)

result = get_sliced_prediction(
    "image.jpg",
    detection_model,
    slice_height=512,
    slice_width=512,
)

Depending on your model requirements, you may need to install additional frameworks alongside SAHI.

Ultralytics (YOLO):

pip install ultralytics>=8.3.161

HuggingFace (Transformers):

pip install transformers>=4.49.0 timm

YOLOv5:

pip install yolov5==7.0.14 sahi==0.12.1

MMDetection (mmdet):

pip install mim
mim install mmdet==3.3.0

Roboflow (Inference):

pip install inference>=0.51.5 rfdetr>=1.6.2

Use YOLO-World for open-vocabulary detection (detecting objects by text description) using model_type="yolo-world".

from sahi import AutoDetectionModel
from sahi.predict import get_sliced_prediction

detection_model = AutoDetectionModel.from_pretrained(
    model_type="yolo-world",
    model_path="yolov8s-worldv2.pt",
    confidence_threshold=0.1,
    device="cuda:0",
)

result = get_sliced_prediction(
    "image.jpg",
    detection_model,
    slice_height=512,
    slice_width=512,
)

SAHI supports three interchangeable backends for postprocessing (NMS, NMM) depending on your hardware and dependencies:

• numpy: Best for CPU-only environments with small/medium prediction counts. No extra dependencies required.
• numba: Best for CPU with large prediction counts. Requires pip install numba. Note: there is a ~1s JIT warmup on the first call.
• torchvision: Best for CUDA GPU availability; fastest for large batches. Requires pip install torch torchvision and a CUDA-enabled environment.

By default, SAHI uses auto-detection, prioritizing torchvision (if CUDA is present), then numba, then numpy.

from sahi.postprocess.backends import get_postprocess_backend

# Check which backend was resolved (triggers auto-detection)
print(get_postprocess_backend())  # Returns "auto" until first postprocessing call