Can SAHI be used with ByteTrack?

[tonyreina]

I've got large images with small objects to detect. I think SAHI might be a good choice for me, but I also want to use ByteTrack (or any object tracking algorithm). Does SAHI work directly with the model.track approach in Ultralytics? From the docs it only appears to work with the bounding box predictions, but not the tracking. Would I just do it in 2 steps then? SAHI predict to saved text files and then ByteTrack on the prediction text files?

[UltralyticsAssistant]

👋 Hello @tonyreina, thank you for your interest in Ultralytics 🚀! We recommend exploring the Docs for detailed guidance, where you can find valuable information on Python and CLI examples. Your use case sounds exciting, and we’re here to help clarify how components like SAHI and ByteTrack can work together.

If this is a ❓ Question about custom workflows or integrations like SAHI and object tracking, we encourage you to share any relevant code snippets, debugging output, or logs to better understand your setup.

If this is a 🐛 Bug Report related to specific model behavior, please provide a minimum reproducible example to help us investigate further.

In the meantime, we encourage upgrading to the latest version of the ultralytics package to ensure you're benefiting from recent enhancements and fixes. You can do this in a Python>=3.8 environment like so:

pip install -U ultralytics

For seamless experimentation, you can leverage any of the following verified environments with preinstalled dependencies:

• Notebooks:
• Google Cloud: See GCP Quickstart Guide
• Amazon AWS: See AWS Quickstart Guide
• Docker: See Docker Quickstart Guide

Join the Community

Join the Ultralytics community where it suits you best:

• Real-time support on Discord 🎧
• Engage in discussions on Discourse
• Share insights and ask questions on our Subreddit

Status

If this badge is green, all Ultralytics CI tests are currently passing, indicating stable performance across all tested YOLO Modes and Tasks.

📌 Please note this is an automated response to assist you promptly. One of our Ultralytics engineers will follow up soon to provide further assistance!

[glenn-jocher]

@tonyreina sAHI currently focuses on improving object detection through sliced inference and does not directly integrate with the model.track approach in Ultralytics or tracking algorithms like ByteTrack. Your suggested two-step process—using SAHI for sliced inference to generate predictions and then applying ByteTrack for tracking on those predictions—would be a practical solution.

[tonyreina]

Thanks @glenn-jocher .

I see the code for batch predictions using Ultralytics and SAHI:

from sahi.predict import predict

predict(
    model_type="yolov8",
    model_path="path/to/yolo11n.pt",
    model_device="cpu",  # or 'cuda:0'
    model_confidence_threshold=0.4,
    source="path/to/dir",
    slice_height=256,
    slice_width=256,
    overlap_height_ratio=0.2,
    overlap_width_ratio=0.2,
)

It appears to be running one image at a time rather than a batch of images. Does the batch here just mean a batch of slices for one image?

Could I specify larger batches of images? My GPU is only being used about 10%.

[glenn-jocher]

@tonyreina you're correct that SAHI processes slices of a single image rather than handling multiple images simultaneously in a batch. Currently, SAHI doesn't support batching multiple images directly. To better utilize your GPU, you could process multiple images concurrently by spawning multiple SAHI processes or threads, but ensure sufficient GPU memory is available.