Github Polytrack ((full)) May 2026
📄 PolyTrack: Fast Multi-Object Tracking with Bounding Polygons
Purpose and motivation
- Accuracy: Polygons match object outlines more closely than boxes, reducing false positives/negatives when shapes vary.
- Robustness: Better handling of partial occlusions and shape deformations.
- Flexibility: Useful for robotics, augmented reality, mapping, and video analytics where precise geometry matters.
for frame in your_video_frames:
masks = your_segmentation_model.predict(frame) # list of binary masks
tracked = tracker.update(masks)
visualize(frame, tracked)
</code></pre>
<h3>Command line</h3>
<pre><code class="language-bash">python track.py --input video.mp4 --model mask_rcnn --output tracks.json
</code></pre>
<h2>📊 Performance</h2>
<p>| Dataset | MOTA | IDF1 | FPS (GPU) |
|---------|------|------|-----------|
| KITTI | 78.4 | 81.2 | 45 |
| BDD100K | 72.1 | 75.8 | 38 |
| YouTube-VIS | 68.3 | 72.5 | 42 |</p>
<h2>🧠 How it works</h2>
<ol>
<li><strong>Predict</strong> – instance segmentation per frame.</li>
<li><strong>Match</strong> – Hungarian algorithm with mask IoU cost matrix.</li>
<li><strong>Filter</strong> – Kalman filter on polygon centroids.</li>
<li><strong>Manage</strong> – birth/death of tracks with memory.</li>
</ol>
<h2>📁 Output format (JSON)</h2>
<pre><code class="language-json">
"tracks": [
To find the project, simply search github polytrack and look for the repository with the most recent commits. (Be wary of abandoned forks; the active branch is usually under a username like polytrack-org or vfx-open-source.) github polytrack
: Saves movement paths and interaction data for ecological analysis. Which one are you looking for? Accuracy: Polygons match object outlines more closely than
