🚗 Nexar Collision Prediction

Predict car collisions from dashcam footage before they happen!

🥉 3rd Place Solution to the Kaggle Nexar Collision Prediction Competition.

📄 Solution write-up: Discussion post on Kaggle.

📦 Installation

1. Create a Python environment:

   conda create -n nexar python=3.10
   conda activate nexar

2. Install Poetry (dependency manager):

   pip install poetry

3. Install the project dependencies:

   poetry install

📁 Data Setup

1. Download the competition dataset from the Kaggle competition page.
2. Place the downloaded ZIP file in the data/raw/ directory:

   data/
   └── raw/
       └── nexar-collision-prediction.zip
   

3. Unzip the file:

   unzip data/raw/nexar-collision-prediction.zip -d data/raw/

🚀 Usage

1. Preprocess the data

Run notebooks/process_data.ipynb. This step extracts and processes features from the raw dashcam footage. It may take a few hours depending on your hardware.

After preprocessing, the data will be organized under data/processed/train/ as follows:

data/processed/train/
└── 00001/                  # Unique video ID
    ├── flows/              # Optical flow tensors
    │   ├── 00.pt
    │   ├── 01.pt
    │   └── 02.pt
    ├── frames/             # Frame tensors (e.g., RGB frames)
    │   ├── 00.pt
    │   ├── 01.pt
    │   └── 02.pt
    └── masks/              # Frame-level vehicle masks
        ├── 00.pt
        ├── 01.pt
        └── 02.pt

Each subfolder contains PyTorch tensors (.pt) representing:

• flows/: Optical flow data between consecutive frames, shape (2, H, W).
• frames/: Processed frame data, shape (3, H, W).
• masks/: Binary masks or target labels for each frame, shape (1, H, W).

2. Train a model

Train a single model on the processed data using notebooks/train_single_model.ipynb. This typically takes 15–30 minutes. Optionally a seed can be set for reproducibility.

🛠️ Project Structure

.
├── data/
│   └── raw/                        # Raw data from Kaggle
│   └── processed/                  # Preprocessed data
├── logs/                           # Logging WandB and checkpoints
├── notebooks/
│   └── process_data.ipynb          # Data preprocessing pipeline
│   └── train_single_model.ipynb    # Model training script
├── src/                            # Source code (data loading, models, utils, etc.)
├── pyproject.toml                  # Poetry config file
└── README.md                       # This file

📫 Contact

For questions or feedback, feel free to reach out via the Kaggle discussion.