- Status
- Research
- Version
- 1.0.0
- Platform
- Desktop app
- License
- MIT
Most public driving datasets are American, European or Chinese, and models trained on them do measurably worse on Indian roads: COCO has no auto-rickshaw class, two-wheelers behave differently and animals appear in traffic. DriveLab is built to measure that gap on your own recordings and then act on it.
How it is put together.
DriveLab is a PyQt6 desktop workbench for a person's own dashcam footage: browsing clips, measuring how off-the-shelf vision models perform on Indian roads, labelling and fine-tuning a detector, fusing GPS with video, and exporting privacy-blurred geotagged imagery. It runs entirely on the user's machine, reads embedded GPS from several dashcam formats, and can push frames to and pull review verdicts from a LabeloxAV server. The README states plainly that a few hours of front-only video is enough to evaluate and fine-tune a small detector, not to train a driving policy.
- 1
Telemetry parsers
Registry of dashcam GPS parsers (Novatek, GPMF, subtitle, sidecar, QuickTime, OSD)
drivelab/core/telemetry/ - 2
Core processing
Import, GPS extraction, events, depth fusion, tracking, training, evaluation, privacy, routes
drivelab/core/ - 3
LabeloxAV connector
Push frames and proposals, pull reviewer verdicts; handoff bundle export
drivelab/core/labelox_api.py - 4
Workspaces
Library, Domain Gap, Annotate, Fine-tune, GPS Fusion, Analytics, Export
drivelab/ui/workspaces/ - 5
UI widgets
Video player, timeline, native QPainter slippy map, hand-painted menus
drivelab/ui/widgets/ - 6
CLI
labelox-check / push / pull commands
drivelab/cli.py
Specification.
- Workspaces
- Library, Domain Gap, Annotate, Fine-tune, GPS Fusion, Analytics, Export
- Models
- YOLO11, RT-DETR, Depth Anything, SegFormer; ByteTrack tracking; CLIP/SigLIP2 semantic frame search
- Dashcam GPS parsers
- Novatek freeGPS (verified on hardware), GoPro GPMF, subtitle NMEA/OSD, sidecar files, QuickTime location (synthetic fixtures)
- Domain gap (200 frames)
- yolo11s (COCO) found 0 auto-rickshaws; a yolo11s fine-tuned on IDD found 83 (unverified detections)
- Monocular distance
- Self-calibrated focal length; depth and width-prior estimates agree at r = 0.893 over 2,694 lead samples; absolute scale unverified (±20%)
- Failure mining
- Detector misses, class switches and flickers found from track continuity with no labelling
- Export
- Frames every N metres, face/plate blur, GPS EXIF, GPX/GeoJSON/CSV, packaging for mapillary_tools
- Requirements
- Linux, Python 3.11+; CUDA GPU optional (CPU inference slow but functional)
- Runs on
- PyQt6 desktop app with a 'drivelab' GUI entry point; GPU optional
Log.
- DriveLab initial releaseDashcam analysis workbench, followed by semantic frame search and a headless CLI.