How to Use SAHI to Detect Small Objects

Detecting small objects is a challenging task in computer vision, yet significant for many use cases. Slicing Aided Hyper Inference (SAHI) is a common method of improving the detection accuracy of small objects, which involves running inference over portions of an image then accumulating the results.

In this guide, we are going to show how to detect small objects with SAHI and supervision. supervision is a Python package with general utilities for use in computer vision projects. The supervision SAHI implementation is model-agnostic, so you can use it across a range of models. We’ll walk through an example of detecting objects on a beach in an image with SAHI.

Here is an example of an image on which inference was run with and without SAHI:

Without further ado, let’s get started!

What is SAHI?

SAHI is a technique for running inference on a computer vision model which aids in detecting smaller objects in an image. SAHI involves running inference on smaller segments of an image, then patching together the results to show all of the detections in an image.

For example, consider a scenario where you are doing an aerial survey and you want to count the number of houses in an image, or a scenario where you are monitoring how busy a beach gets at peak times. In both cases, predictions may be small, and SAHI can help boost the detection rate of objects of interest.

The approach that we walk through for using SAHI is model-agnostic, with support for all models with supervision data loaders. Whether you are using Ultralytics YOLOv8, a model on Roboflow, a vision model in the Transformers Python package, or another model with a supported data loader, you can follow our instructions below to use SAHI.

Step #1: Set Up a Model

First, we will need a model on which to run inference. For this guide, we will be using the Ultralytics YOLOv8x model. You can use any Roboflow model on Universe, private models associated with your account, or other model types with supervision data loaders.

With a model ready, we can start writing a script to use SAHI. Create a new Python file and add in the following code:

import supervision as sv
import numpy as np
import cv2
from ultralytics import YOLO

model = YOLO("yolov8x.pt")

def callback(x: np.ndarray) -> sv.Detections:
    result = model(x, verbose=False, conf=0.25)[0]
    return sv.Detections.from_ultralytics(result)

In this code, we define a callback function that, given an image, will run inference using the Ultralytics package and return the results. When you first run the program, you will be asked to login to Roboflow with an interactive login interface.

Step #2: Run Inference with the SAHI Slicer

Now that we have a model set up, we can run inference with the SAHI slicer. We can do so using the following code:

image = cv2.imread("./image.jpg")

slicer = sv.InferenceSlicer(callback=callback)
sliced_detections = slicer(image=image)

This code creates a slicer object which accepts the model inference callback function we defined in the previous step. We then use the slicer object to run inference on an image. The slicer object splits up the specified image into a series of smaller images on which inference is run.

To display SAHI predictions, we can use the following code:

box_annotator = sv.BoxAnnotator()

sliced_image = box_annotator.annotate(image.copy(), detections=sliced_detections)

sv.plot_image(sliced_image)

Here are the results from inference with and without SAHI:

Above, inference with SAHI was able to capture significantly more objects than inference without SAHI.

Conclusion

SAHI is a technique for improving the detection rate for smaller objects with object detection models. SAHI works by running inference on “slices’ of an image, then combining the results across the whole image. supervision has a built-in SAHI feature you can use to run SAHI on any model for which there is a data loader in supervision.

Combined with the Roboflow API or the self-hosted Roboflow Inference solution, you can use the SAHI technique to run inference on models hosted on Roboflow.

In this guide, we demonstrated how to detect small objects with SAHI and supervision. Now you have all of the resources you need to use SAHI to improve the detection rate of smaller objects using object detection models.