🔥 FlashBack — Rewinding Fire to Its Origin

⏪ Physics-Aware Fire Origin Tracing with NVIDIA Cosmos-Reason2

An AI system that physically traces fire back to its origin from surveillance footage,
powered by NVIDIA Cosmos-Reason2's physical reasoning to understand combustion, convection, and smoke dynamics.

While conventional fire detection asks "Is there a fire?",
FlashBack answers "Where did it start, and how did it spread?" 🎯

🎬 Demo

demo_2b.mp4

⚙️ How It Works

FlashBack operates in three stages, leveraging Cosmos-Reason2 as a physics-aware video reasoning engine:

flowchart LR
    A["🎥 Surveillance<br/>Video"] -->|"fps=1<br/>sampling"| B["🧠 Cosmos-Reason2<br/>Physics-Aware<br/>Reasoning"]
    B -->|"(x, y)<br/>coordinates"| C["📍 Origin Point<br/>+ Spread Direction"]
    B --> D["💭 Chain-of-Thought<br/>Combustion Physics<br/>Convection Patterns<br/>Smoke Dynamics"]
    C --> E["👁️ Lucas-Kanade<br/>Optical Flow<br/>Tracking"]
    E --> F["📊 Output<br/>Video + Image<br/>Dashboard"]

    style A fill:#2d2d2d,stroke:#76B900,color:#fff
    style B fill:#76B900,stroke:#333,color:#fff
    style C fill:#FF6F00,stroke:#333,color:#fff
    style D fill:#1a1a2e,stroke:#76B900,color:#ccc
    style E fill:#2d2d2d,stroke:#76B900,color:#fff
    style F fill:#2d2d2d,stroke:#FF6F00,color:#fff

Loading

🔬 Stage 1 — Physics-Aware Prompting

Cosmos-Reason2 is prompted as a fire physics expert. The model analyzes video frame-by-frame, performing Chain-of-Thought reasoning grounded in combustion physics.

"The pattern of flames spreading outward and smoke accumulating above indicates a typical growth-phase fire. The origin is at the lower section where fuel sources are concentrated, spreading upward via thermal convection." — Cosmos-Reason2 CoT reasoning example

📍 Stage 2 — Coordinate-Based Origin Tracing

The model outputs the fire origin as normalized coordinates (x, y). It simultaneously infers both a textual description (e.g., "lower-left of the greenhouse") and precise coordinates (0.25, 0.75), which are overlaid on the original frame as an origin marker + spread direction arrows.

👁️ Stage 3 — Optical Flow Tracking

Lucas-Kanade optical flow tracks the origin coordinates across all frames. Even with camera movement, the origin marker stays locked to the correct position, enabling real-time visualization of fire propagation in the demo video.

🧪 Physics Reasoning Principles

Principle	What the Model Analyzes	How It Traces the Origin
🌡️ Convection	Rising hot air currents, smoke transport paths	Below smoke accumulation point = origin
🔥 Combustion	Fuel density vs. flame intensity	Identify initial fuel concentration
💨 Propagation	Heat-transfer-driven outward spread	Reverse-trace from the spread center
🌫️ Smoke Color	Smoke color → burning material type	Infer origin location + fuel source
⏱️ Temporal	Ignition → Growth → Flashover → Decay	Reverse direction from frame changes

📈 Results

Cosmos-Reason2-2B / 8B  |  11 scenes  |  FLAME 6 / SMOKE 2 / NORMAL 3

Metric	Score
🎯 Fire Origin Tracing	100%
⏱️ Temporal Reasoning	100%
↗️ Spread Direction	85.7%

🖼️ Origin Visualization

🔥 sample1 — FLAME · Fire origin + spread direction in greenhouse

🌫️ sample2 — SMOKE · Reverse-tracing origin from smoke dispersion

🚀 Quick Start

git clone https://github.com/Hann1n/FlashBack.git
cd FlashBack

python -m venv .venv && source .venv/Scripts/activate  # Windows
# source .venv/bin/activate  # Linux/macOS
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu124
pip install -r requirements.txt

# 🔄 Full pipeline
python run.py

# Or step by step
python -m src.core.detection      # 🧠 Cosmos-Reason2 inference (2B/8B, GPU)
python -m src.core.visualize      # 🖼️ Origin overlay images
python -m src.ui.dashboard        # 📊 HTML dashboard
python -m src.core.video           # 🎬 Demo video with optical flow
streamlit run src/ui/app.py       # 🌐 Interactive dashboard

📁 Project Structure

FlashBack/
├── 🚀 run.py                      # Pipeline entry point
├── 📂 src/
│   ├── config.py                  # Centralized path configuration
│   ├── 🧠 core/                   # Inference & logic
│   │   ├── inference.py           # Cosmos-Reason2 model wrapper
│   │   ├── detection.py           # Fire detection + origin inference
│   │   ├── new_data.py            # New dataset inference pipeline
│   │   ├── visualize.py           # Origin marker visualization
│   │   └── video.py               # Demo video (optical flow tracking)
│   ├── 🖥️ ui/                     # Dashboards & visualization
│   │   ├── app.py                 # Streamlit interactive dashboard
│   │   ├── dashboard.py           # Standalone HTML dashboard
│   │   └── fiftyone_builder.py    # FiftyOne dataset builder
│   └── 🔧 utils/                  # Shared utilities
│       └── common.py              # imread_unicode, origin helpers
├── 📂 data/                       # Local datasets (download separately)
├── 📂 reports/                    # Inference results (JSON)
├── 🎬 demo/                       # Demo video
├── 🖼️ assets/                     # README images
├── requirements.txt
└── README.md

🛠️ Technical Stack

Component	Technology
🧠 Model	Cosmos-Reason2-2B / 8B (Qwen3VL)
👁️ Tracking	Lucas-Kanade Optical Flow
🎥 Video	PyAV backend (Windows FFmpeg workaround)
⚡ Inference	fps=1, temp=0.6, CoT reasoning enabled
📊 Visualization	OpenCV, Plotly, Chart.js, Streamlit

Built with 💚 for NVIDIA Cosmos Cookoff 2026
Uses Cosmos-Reason2 under NVIDIA Open Model License