LandMark is a large-scale 3D city scene modeling and rendering system designed for researchers and developers working with extensive real-world data. It extends GridNeRF to enable efficient training and high-resolution rendering (up to 4K) of city-scale environments, supporting novel view synthesis and scene manipulation like object removal or addition.

How It Works

LandMark builds upon GridNeRF, introducing significant optimizations for training and rendering efficiency through parallelism, custom operators, and kernels. It supports various parallel training strategies (Sequential, Branch, Plane, Channel, Hybrid) and distributed rendering, allowing it to handle over 100 square kilometers of city data with billions of learnable parameters. This approach facilitates the creation of detailed, large-scale 3D neural scenes.

Quick Start & Requirements

• Install: Clone the repository, set PYTHONPATH, create a Conda environment (python=3.9.16), and install dependencies (pip install -r requirements.txt).
• Prerequisites: NVIDIA GPU with CUDA 11.6, PyTorch 1.13.1.
• Dataset: Requires datasets formatted with images/ and transforms_train.json/transforms_test.json, ideally 250-300 images with sufficient overlap. The MatrixCity dataset is recommended.
• Documentation: DocumentationSite

Highlighted Details

• Achieves efficient training of 3D neural scenes on over 100 sq km of city data.
• Supports novel view rendering at 4K resolution with over 200 billion parameters.
• Enables feature extensions like layout adjustment (adding/removing buildings) and scene stylization (lighting, seasons).
• Offers a real-time distributed rendering system capable of over 30 FPS for large scenes.

Maintenance & Community

• Developed by the LandMark Team at Shanghai AI Laboratory.
• Community support and contributions are encouraged.

Licensing & Compatibility

• The repository does not explicitly state a license in the README. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

• Native datasets used in examples are not released due to confidentiality.
• Requires significant computational resources (multiple A100 GPUs recommended for optimal performance).