A PyTorch-based toolbox for LiDAR-based 3D object detection, supporting multiple state-of-the-art models and datasets.
OpenPCDet is an open-source PyTorch toolbox designed for LiDAR-based 3D object detection. It provides a unified implementation of numerous state-of-the-art detection models, enabling researchers and developers to train and evaluate models on major autonomous driving datasets. The project solves the problem of fragmented codebases by offering a single, well-structured framework that supports both one-stage and two-stage detection approaches.
Researchers and engineers working on 3D perception for autonomous vehicles, robotics, and computer vision who need a reliable codebase to experiment with or deploy LiDAR-based detection models.
Developers choose OpenPCDet for its comprehensive model zoo, consistent code quality, and active updates with cutting-edge methods. Its modular design and support for multiple datasets reduce implementation overhead, allowing users to focus on innovation rather than reinfrastructure.
OpenPCDet Toolbox for LiDAR-based 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.
Implements a wide range of state-of-the-art 3D detection models like PointRCNN, PV-RCNN, and Voxel R-CNN, with pretrained weights available for multiple datasets, as shown in the detailed model zoo tables.
Supports key autonomous driving datasets including KITTI, Waymo, NuScenes, ONCE, and Argoverse2, enabling easy benchmarking and transfer learning across domains.
Features data-model separation and a consistent 3D box definition (x, y, z, dx, dy, dz, heading), simplifying extension to custom datasets, as illustrated in the design pattern diagrams.
Regularly integrates cutting-edge methods, such as DSVT for real-time inference and BEVFusion for multi-modal detection, with changelog updates through 2023.
Setting up dependencies like specific PyTorch and spconv versions can be tricky, with the README noting compatibility ranges (e.g., PyTorch 1.1~1.10, spconv 1.0~2.x) that may require manual troubleshooting.
While extensible, adding new models or modifying architectures demands deep familiarity with the codebase structure, despite the provided custom dataset tutorial.
Training requires multiple GPUs and long hours (e.g., 5 hours for PV-RCNN on KITTI with 8 GPUs), making it inaccessible for teams with limited computational resources.
Newer additions like multi-modal fusion or temporal models have guidelines but may lack comprehensive examples, leaving users to rely on community contributions or source code diving.