Can slice detection and tracking of rotating frames be achieved at the same time?

[ZSBSB]

Search before asking

• I have searched the Supervision issues and found no similar feature requests.

Question

Thank you very much for your work, I have a question, currently the supervision library can implement slice detection and tracking of rotated frames separately, so when it is necessary to track a small target in the image, is it possible to use slice detection to get the small target first, and then track it later, and finally output the result? Looking forward to your replies.

Additional

No response

[JESUSROYETH]

Hi! Yes, it's definitely possible to perform both slicing-based detection and frame-by-frame tracking (even with rotated objects) using the supervision library. In fact, the workflow you're describing is a great approach for handling small object tracking.

🔍 Detection + Tracking: How it Works Together

One of the nice things about supervision is how modular it is — you can mix and match components to build a pipeline that fits your needs. Here’s how to tackle this:

---

🧩 Step 1: Detect Small Objects with Slicing

When working with tiny objects that might be missed in full-frame inference, you can use InferenceSlicer. It splits the image into smaller patches, runs detection on each one, and then stitches the results back together. This gives you much better accuracy on small targets.

It returns an sv.Detections object with all your bounding boxes and confidences.

---

🧠 Step 2: Track Across Frames

Once you have detections, pass them to a tracker like ByteTrack, which is also integrated into supervision. The tracker assigns a unique ID to each detected object and keeps track of them across frames.

---

🔄 Step 3: Handle Rotated Objects

If your objects are tilted or not aligned with the image axes, no problem — supervision also supports rotated bounding boxes and can calculate IoU between them. That helps improve the accuracy of detection merging and tracking in more complex scenarios.

---

✅ Suggested Pipeline

Here's a high-level view of how your code would look:

1. Load your object detection model (e.g. YOLO).
2. Initialize an InferenceSlicer with your model.
3. For each video frame:
   • Run the slicer → get sv.Detections.
   • Pass the detections to the tracker (e.g. ByteTrack).
   • The tracker will assign and update tracker_ids.
4. Visualize: Use supervision's annotators to draw boxes, tracker IDs, and more.

---

Hope this helps!