ISOFRAMES
GitHub

DashLab

Train object detectors and visual models on dashcam footage without labelling any of it

Source on GitHub All systems
Status
Live
Version
0.1.0
Platform
GPU workstation
Source
github.com/Sherin-SEF-AI/DashLab
License
MIT
01 · Problem

Labelling dashcam footage by hand is the bottleneck for training detectors and visual models on it. DashLab derives supervision from the footage itself (teacher predictions, self-distillation, OSD telemetry) and reports scores as agreement with the teacher, not ground truth.

02 · Architecture

How it is put together.

DashLab produces three models from raw dashcam video with no human annotation: a YOLO11n detector distilled from a YOLO11x teacher, a DINO-style self-supervised ResNet-18 backbone, and an embedding-density anomaly scorer. It also reads speed, GPS and clock off the burned-in OSD, supports query-by-example and CLIP text search, and clusters braking events by GPS. A PyQt6 desktop app with eight workspaces runs the 31 pipeline stages as subprocesses and browses the results. It was built and measured on a Jetson AGX Orin using 128 clips (22,805 sampled frames) of 4K30 footage.

Frame extractionpipeline/01_extract.pySelf-supervisionpipeline/Student training and ev…pipeline/08_train_student.pyAnalysispipeline/Export and livepipeline/15_export.pyDesktop appgui/Frame extractionpipeline/01_extract.pySelf-supervisionpipeline/Student training and ev…pipeline/08_train_student.pyAnalysispipeline/Export and livepipeline/15_export.pyDesktop appgui/
FIG. 01Modules and data flows, from the repository layout and documentation.
  1. 1

    Frame extraction

    Decode clips, crop hood and OSD, sample frames incrementally

    pipeline/01_extract.py
  2. 2

    Self-supervision

    DINO backbone training, OSD OCR, teacher pseudo-labels, embeddings

    pipeline/
  3. 3

    Student training and evaluation

    Train student detector, self-training, compare checkpoints, ground-truth sampling and eval

    pipeline/08_train_student.py
  4. 4

    Analysis

    Anomaly scoring, retrieval, GPS hotspots, class discovery, tracking, TTC, CLIP search

    pipeline/
  5. 5

    Export and live

    ONNX / TensorRT export and benchmark; real-time camera or video loop

    pipeline/15_export.py
  6. 6

    Desktop app

    PyQt6 app running stages as subprocesses and browsing results

    gui/
03 · Capabilities

Specification.

Models
YOLO11n student detector (11 classes), self-supervised ResNet-18 backbone (512-d), anomaly / novelty scorer
Detector (vs teacher, 25 held-out clips, 1280px)
mAP50 0.583 / mAP50-95 0.482 over 11 classes; agreement with teacher, not ground truth
Class discovery
Clustering in the SSL embedding space separated autorickshaw, minibus and goods_carrier (~33,000 boxes relabelled)
OSD telemetry
Speed, GPS and clock read from burned-in overlay; 99.98% per-digit accuracy over 250,980 digits
Deployment (Jetson AGX Orin)
TensorRT FP16 57.1 fps; PyTorch GPU 35.1 fps; live loop 24.7 fps on 1080p at 960 width
Search
Query-by-example and CLIP text search; zero-shot scene tags (day/night 99.0% on 14,810 frames)
Dataset
DashLab Traffic v1.1: 22,805 frames, 296,616 pseudo-label boxes, 11 classes, anonymised
Desktop app
PyQt6, eight workspaces, stages as subprocesses with progress, ETA and live GPU/CPU/RAM telemetry
Runs on
PyQt6 desktop front end, but training and inference need CUDA/TensorRT; built and measured on a Jetson AGX Orin
Stack
  • Python
  • PyTorch
  • Ultralytics (YOLO11)
  • PyQt6
  • OpenCV
  • scikit-learn
  • SciPy
  • pandas
  • ONNX / ONNX Runtime
  • TensorRT
  • CLIP
04 · Changelog

Releases.

  1. v0.1.0
    Label-free training pipeline and desktop app for dashcam footage.
05 · Related

More systems.

Perception and ADAS · LiveLyne ADAS

Real-time, on-device driver assistance for any Android phone.

Perception and ADAS · ResearchDRIVYX

Perception and waypoint control training platform for the India Driving Dataset (IDD)

Next systemDriveLab

A desktop application for working with consumer dashcam footage