An open-source PyTorch toolbox for general 3D object detection, supporting LiDAR, camera, and multi-modal models.
MMDetection3D is an open-source toolbox for general 3D object detection, built on PyTorch as part of the OpenMMLab project. It provides a unified platform for training, evaluating, and deploying state-of-the-art 3D detection models that work with LiDAR, camera, or multi-modal data. The toolbox solves the problem of fragmented implementations by offering a comprehensive, efficient codebase that supports a wide range of datasets and model architectures for applications like autonomous driving and robotics.
Researchers and engineers working on 3D perception, particularly in fields like autonomous vehicles, robotics, and augmented reality, who need a flexible and efficient framework for developing and benchmarking detection models. It is also suitable for academic institutions and industry teams leveraging OpenMMLab's ecosystem for computer vision projects.
Developers choose MMDetection3D for its extensive model zoo, high training efficiency, and seamless integration with 2D detection tools from the OpenMMLab suite. Its modular design and support for multiple sensors and datasets make it a versatile, production-ready platform that accelerates research and deployment compared to building custom solutions or using less comprehensive alternatives.
OpenMMLab's next-generation platform for general 3D object detection.
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.
Includes over 500 models from MMDetection and numerous 3D architectures like VoteNet and PointPillars, offering state-of-the-art implementations for various tasks.
Benchmarks show faster training speeds compared to other codebases, such as 358 samples/second for VoteNet vs. 77 in votenet, optimizing development cycles.
Provides out-of-the-box implementations for LiDAR, camera, and fused sensor data, including models like MVXNet and BEVFusion for robust perception.
Naturally integrates with MMDetection's 300+ models, enabling hybrid 2D-3D workflows and component reuse, as highlighted in the README.
Directly supports major indoor and outdoor datasets like KITTI, nuScenes, Waymo, and ScanNet, facilitating comprehensive benchmarking.
Installation requires specific PyTorch versions and OpenMMLab dependencies, and dataset preparation involves detailed steps that can be time-consuming.
Tight coupling with the OpenMMLab framework means adopting its tools and conventions, which may limit flexibility for teams using other ecosystems.
The modular design and extensive configuration options, while powerful, require significant upfront learning, especially for those new to 3D perception.
Regular updates, such as v1.4.0 refactoring the Waymo dataset, can introduce compatibility issues and require code adjustments.