Obscuro

Desktop-first tooling for anonymizing dashcam footage: optimized ONNX/SAHI detector pipeline, FastAPI backend, CLI utilities, and an Electron UI with live previews. Trackers include ByteTrack/BotSort/OC-SORT plus a fused distance+shape+embedding tracker (with Hybrid SOT built on top for detector gaps). Detection targets are configurable—blur license plates and heads by default, or extend to other classes when using segmentation-enabled models.

Documentation

• Overview & architecture
• FastAPI endpoints
• CLI commands
• Python API guide
• Configuration reference

Quick start

Example of the Electron desktop frontend with live video preview and anonymization controls

# Create the dev environment (Python 3.12+)
uv sync

# Run the backend API
uv run blur-api

# Launch the Electron desktop frontend (in src/frontend/)
npm install
npm run dev

Need CUDA? Only enable the GPU dependencies on NVIDIA-enabled hosts:

uv sync --extra gpu

Container images (CPU and GPU) can be built with docker build -f Dockerfile[.gpu] . or consumed from GitHub Container Registry. See the docs linked above for volumes, ports, and auto-start configuration. Visual tracking backends like TrackerNano require user-provided weights (see docs); they are not bundled in the repo.

Contributing

Open issues and pull requests on GitHub. Run uv run pytest and uv run ruff check locally before submitting. Artwork and screenshots live under docs/assets/.

License & attribution

Licensed under AGPL-3.0-or-later.

• Detector checkpoints are trained from Ultralytics YOLO11 pretrained weights (Ultralytics code is AGPL-3.0: https://github.com/ultralytics/ultralytics/blob/main/LICENSE).
• The MobileNetV3 embedding ONNX (models/tracking/mobilenet_v3_small_embed.onnx) is exported from torchvision.models.mobilenet_v3_small with the classifier stripped (torchvision is BSD-3-Clause: https://github.com/pytorch/vision/blob/main/LICENSE).
• See LICENSE and the docs for full attribution guidance.