DriveStudio is a codebase for omni urban scene reconstruction and simulation. It enables the joint reconstruction of backgrounds, vehicles, and non-rigid objects from driving data. The framework supports multiple Gaussian representations, allowing for flexible modeling of various scene elements. The core problem addressed is to accurately and robustly reconstruct 3D urban environments from driving sequences.
DriveStudio distinguishes itself through its unified data system supporting diverse datasets, flexible multi-camera training options, and integration with gsplat for advanced rasterization. It also includes features like camera pose refinement and ground truth bounding box correction, enhancing reconstruction accuracy. The acceptance of OmniRe by ICLR 2025 as a spotlight paper highlights its innovative approach.
- Multi-Camera Support: Flexible configurations for single, multiple, or all cameras during training.
- Dataset Integration: Supports popular datasets like Waymo, NuScenes, and KITTI with detailed processing instructions.
- gSplat Integration: Leverages gsplat for advanced rasterization techniques like anti-aliasing.
- Pose Refinement: Provides a mechanism to refine camera poses for improved accuracy.
- Robustness: Implements features to address noisy ground truth boxes and camera issues.
- Extensible: Configurable through YAML files, allowing users to easily experiment with different settings and configurations.
- Flexible Training: Supports both single-representation and multi-representation Gaussian trainers.
DriveStudio is an actively developed project with recent updates, including acceptance by ICLR 2025. The codebase includes detailed documentation & instructions for data processing and usage. The repository has a good level of community engagement with regularly addressing issues and incorporating new features. Frequent commits indicate ongoing development and maintenance.
DriveStudio is beneficial for researchers and developers working on 3D urban scene understanding, autonomous driving, and simulation. It facilitates research and development by providing a robust, flexible, and well-documented framework to reconstruct realistic 3D environments. It simplifies the process of working with various driving datasets and exploring advanced scene reconstruction techniques.