See the model think.
A live walkthrough of the perception layer that powers the connected ecosystem — real YOLOv8 + ByteTrack detections on Caltrans CCTV footage, rendered frame-accurate in your browser. Scrub, filter by confidence, and watch the ground-truth tally build.
Live tally
Unique tracks counted so far in this clip
Full-session ground truth
Corrected export across the complete test1.mp4 session
Source clip: test1.mp4 (1280×720) · detections precomputed with the desktop tool, replayed here as a compact JSON overlay.
From pixels to ground truth
Five stages turn raw camera feeds into validated, training-ready datasets.
Ingest
Caltrans CCTV clips or live RTSP / webcam streams
Detect
YOLOv8 inference, auto CUDA / Apple MPS / CPU
Track
ByteTrack assigns persistent IDs across frames
Correct
Frame-by-frame human edits, non-destructive
Export
9 formats — COCO, YOLO, CVAT, MOT, CSV, MP4…
One desktop tool, nine workspaces
Native PySide6 / Qt app for macOS & Windows. v2.2.0 ships hybrid GPU packaging — a universal installer that auto-fetches the matching CUDA build at first run.
Preprocessing
Single-video detect + track with ROI drawing and live ETA.
Batch
Folder-tree scheduler, 1–100 parallel workers, atomic cancel.
Correction
Track editor with keyboard shortcuts and 9 export formats.
Performance
Per-session stats, A/B model compare, confusion matrix.
Analytics
Heatmaps, OD matrix, speed + direction, HTML report.
Insights
Anomaly detection and dataset-health scoring.
Training
Active-learning queue built from unused corrections.
Models
Import, download, rename, and A/B-compare weights.
Live
Webcam / RTSP A/B detection with per-ROI alerts.
Model performance
5×5 confusion matrix exported from the Performance tab — the basis for vendor CV benchmarking.
Export formats
The perception layer of a connected corridor
Detected positions, speeds, and counts feed the V2X applications and deployment roadmap downstream.