Implementation:Obss Sahi Get Prediction

Knowledge Sources	SAHI SAHI Docs
Domains	Object_Detection, Computer_Vision, Inference
Last Updated	2026-02-08 12:00 GMT

Overview

Concrete tool for running single-image object detection inference with coordinate remapping provided by the SAHI library.

Description

get_prediction() performs object detection on a single image (or image slice) and returns a PredictionResult containing a list of ObjectPrediction objects. It orchestrates three internal steps:

Read the image as a PIL Image and convert to a contiguous numpy array
Call detection_model.perform_inference() to run the forward pass
Call detection_model.convert_original_predictions() with shift_amount and full_shape to remap coordinates to full-image space

The function also applies optional class filtering (exclude_classes_by_name/exclude_classes_by_id) and an optional postprocessing step. It tracks timing for profiling via durations_in_seconds.

Usage

Use this function for running detection on individual image slices within the sliced inference loop. It is called once per slice by get_sliced_prediction(). Can also be used standalone for single-image inference without slicing.

Code Reference

Source Location

Repository: sahi
File: sahi/predict.py
Lines: L56-135

Signature

def get_prediction(
    image,
    detection_model,
    shift_amount: list | None = None,
    full_shape=None,
    postprocess: PostprocessPredictions | None = None,
    verbose: int = 0,
    exclude_classes_by_name: list[str] | None = None,
    exclude_classes_by_id: list[int] | None = None,
) -> PredictionResult:
    """Perform prediction on a single image.

    Args:
        image: str or np.ndarray - image path or numpy array
        detection_model: DetectionModel instance
        shift_amount: [shift_x, shift_y] to remap slice to full image
        full_shape: [height, width] of the full image
        postprocess: Optional PostprocessPredictions for single-image NMS
        verbose: 0=silent, 1=print timing
        exclude_classes_by_name: Class names to exclude
        exclude_classes_by_id: Class IDs to exclude
    """

Import

from sahi.predict import get_prediction

I/O Contract

Inputs

Name	Type	Required	Description
image	str or np.ndarray	Yes	Image path or numpy array to detect on
detection_model	DetectionModel	Yes	Initialized detection model from AutoDetectionModel
shift_amount	list[int]	No	[shift_x, shift_y] offset for coordinate remapping (default [0,0])
full_shape	list[int]	No	[height, width] of original image (inferred if None)
postprocess	PostprocessPredictions	No	Optional single-image postprocessing
verbose	int	No	Verbosity level (0=silent, 1=print timing)
exclude_classes_by_name	list[str]	No	Class names to filter out
exclude_classes_by_id	list[int]	No	Class IDs to filter out

Outputs

Name	Type	Description
return	PredictionResult	Contains .object_prediction_list (list of ObjectPrediction with remapped coords) and .durations_in_seconds (dict with "prediction" and "postprocess" timing)

Usage Examples

Single Slice Inference

from sahi import AutoDetectionModel
from sahi.predict import get_prediction

# Initialize model
model = AutoDetectionModel.from_pretrained(
    model_type="ultralytics",
    model_path="yolov8n.pt",
    device="cuda:0",
)

# Run detection on a slice with coordinate remapping
result = get_prediction(
    image=slice_numpy_array,  # numpy array from slice_image()
    detection_model=model,
    shift_amount=[120, 240],  # slice offset in full image
    full_shape=[1920, 2560],  # original image [height, width]
)

for pred in result.object_prediction_list:
    print(f"Class: {pred.category.name}, Score: {pred.score.value:.2f}")
    print(f"BBox (full image coords): {pred.bbox.to_xyxy()}")

Standalone Full-Image Inference

from sahi.predict import get_prediction

result = get_prediction(
    image="path/to/image.jpg",
    detection_model=model,
    # No shift_amount needed for full image
)
print(f"Found {len(result.object_prediction_list)} objects")

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment