Plate Tracker
A YOLOv8 license-plate detector's raw per-frame output, replayed as a clickable overlay in sync with the source video — no live inference, no backend.
Introduction
Plate Tracker shows what a YOLOv8 license-plate detector actually found in a real test clip — every box, every confidence score, every frame — without running the model live for each visitor. Hover or focus any box in the demo below to see the exact coordinates and confidence behind it.
The detector runs once, offline. What ships to the browser is its recorded output.
Context
Demoing a detector live in a browser means paying inference cost on every visit and keeping a model server up just to show a portfolio piece.
Run YOLOv8 with tracking once, offline, over a real test clip; export every frame's boxes, confidence and track id to JSON; replay that JSON client-side, synced to the looping video by timestamp.
Methodology
YOLOv8 with ByteTrack processes the clip frame by frame at 30fps, and every frame's boxes, confidence scores and track ids are exported to a single JSON file. The browser's only job is to look up the frame closest to the video's current playback time and draw whatever was recorded there — the same technique subtitle files use to stay in sync with a video, applied to bounding boxes instead of text.
- frame_index / timestamp_seconds
- Position in the clip
- bbox_xyxy
- Corner coordinates in source pixels
- confidence
- Detector's score for that box
- id
- ByteTrack's identity for the object, or null
Interactive
Results
- 1,800 frames at 30fps processed once offline; 83.7% contain at least one detection.
- 2,052 individual detections replayed in the browser, confidence 0.25–0.84 (avg 0.54).
- The entire interactive demo is a video file plus a 760KB JSON — no backend, no per-visit inference cost.
Technical explanation
Coordinates are stored in the source video's native pixel space (848×478), not normalized — the overlay is an SVG sized to that exact viewBox and stretched over the video element, so a box drawn at pixel (292, 248) lands on the video's pixel (292, 248) regardless of how large the player is rendered on screen.
Because the lookup runs on every animation frame against the video's own currentTime, the overlay stays correct whether the video is autoplaying, paused, or scrubbed by hand — there is no dependency on playback events firing at any particular rate.
Conclusion
- Replaying a recorded inference pass is a legitimate way to demo a model — it removes GPU cost and cold-start latency from the demo entirely, at the cost of it being one fixed clip rather than anything the visitor can upload.
- The tracker occasionally assigns two ids to the same physical plate for a few frames — a real limitation of the tracking step, left in rather than cleaned up, because it is honest about what the model does.
- Storing coordinates in native pixels instead of normalizing them upfront moved the scaling math to one place (the overlay) instead of scattering it through the export script.