Autonomous driving

LabeloxAV

A data engine for autonomous driving, built for Indian roads. Raw fleet footage in, governed labels and retrained models out, with a confidence gate between the machine and the ground truth.

  • Python
  • Docker
  • Active learning
  • HD maps
  • 178 classes
01 / The problem

The classes global datasets never saw

Open autonomous-driving datasets were collected on roads that are lane-marked, largely rule-following, and populated by a short list of vehicle types. An ontology trained on them has no category for an autorickshaw, no prior for cattle on the carriageway, and nothing sensible to say about an overloaded two-wheeler, a hand cart, or a pothole deep enough to matter.

The usual answer is to label more data. That runs into the second problem: at fleet scale, most footage is redundant. Labelling it uniformly spends the budget on the ordinary and starves the rare-and-risky moments that actually move the model.

02 / How it works

A closed loop, not a labelling queue

INGESTraw fleet footageFUSEthree-path auto-labelGATEcalibrated confidenceREVIEWhuman, governedMINErare + riskyRETRAINclosed loopOne ontology · 178 governed classes · per-object dynamics · HD maps

Machine labels come from a three-path fusion pipeline rather than a single detector. Every label then passes a calibrated confidence gate that decides what a human ever sees: high-confidence labels flow through, uncertain ones are routed to review. Active learning mines the rare and risky moments out of the remaining footage, and the models retrain on what came back. The loop is governed: one ontology, versioned classes, and an audit trail behind every export.

LabeloxAV annotation console: an Indian street scene with detected vehicles, pedestrians and autorickshaws boxed, alongside the class panel
The annotation console on real footage. Mixed traffic, roadside stalls and pedestrians in the carriageway, with the frame strip and class panel that reviewers work through.
03 / What's proven

Install and constraints

PropertyDetail
InstallOne command on any machine with Docker: generates secrets, migrates the schema, seeds the ontology, creates the first admin, prints the token
OntologyOne ontology, 178 governed classes
No GPU neededAnnotation, review, governance, export and search all work without CUDA
Failure modeThe model paths that need CUDA refuse rather than fabricate
ConsoleWeb console at localhost:3000
DeploymentGPU, TLS and backups documented separately
04 / Related

The rest of the pipeline

  • DriveLab: desktop workbench for dashcam footage: GPS telemetry extraction, domain-gap evaluation on Indian roads, annotation and fine-tuning, privacy-blurred Mapillary export. Part of LabeloxAV.
  • RabbitAV: the on-device end: a windshield phone doing forward collision warning and IMU-plus-vision road hazard mapping, floor device class 3-4 GB RAM with CPU-only INT8 inference.
  • Orvex: the ingest side: telemetry extraction, calibration, synchronization and dataset assembly from GoPro, Insta360 and Android recordings.