Research and development for ML-based wildfire smoke detection at Pyronear.
Pyronear deploys fixed 360° cameras on antenna towers that capture images every 30 seconds. On each Raspberry Pi, a YOLO model detects smoke candidates. Full frames and bounding box info are sent to a server where a second-stage temporal model verifies detections — reducing false positives while maintaining high recall.
vision-rd/
├── lib/ # Shared packages used across experiments
│ └── pyrocore/ # Types, protocols, and base model (TemporalModel ABC)
├── literature_survey/ # Paper collection, notes, and thematic summary
├── experiments/ # R&D experiments (each a self-contained uv project)
│ ├── template/ # Starter project — copy to begin a new experiment
│ ├── README.md # How to create and manage experiments
│ └── GUIDELINES.md # Standards: uv, ruff, DVC, reproducibility
├── CONTRIBUTING.md # How to contribute
├── README.md # This file
└── LICENSE # Apache 2.0
- 📚 Literature Survey — 28 papers on temporal models, video foundations, smoke detection, and related topics
- 🧪 Experiments README — How to start a new experiment
- 📏 R&D Guidelines — Standards for reproducibility, tooling, and benchmarking
See CONTRIBUTING.md for how to get involved — adding papers, running experiments, and code style conventions.