OpenPCDet

OpenPCDet is a Python toolkit for training and testing AI models that detect objects in 3D space using LiDAR sensors. LiDAR is the kind of laser-based sensor that lets self-driving cars see the world around them, producing clouds of 3D data points instead of flat camera images. This project focuses on turning that raw point-cloud data into labeled detections: cars, pedestrians, cyclists, and similar objects placed in 3D space.

The project was built by the OpenMMLab research group and serves as the official release of several published research methods, including PointRCNN, PV-RCNN, and MPPNet. Each of those is a different algorithm for processing point clouds and placing bounding boxes around detected objects. The codebase is structured so researchers can swap in different detection methods and test them against common benchmarks without rewriting the whole pipeline.

Supported datasets include KITTI (a long-standing benchmark for autonomous driving research), Waymo Open Dataset (a large-scale dataset from Waymo's self-driving project), and NuScenes (a dataset that includes camera and radar data alongside LiDAR). The toolkit handles both single-frame detection and multi-frame detection, where the model uses a short history of frames to improve accuracy.

To use it, you set up a Python environment with PyTorch, install the required dependencies, download one of the supported datasets, and then train or evaluate a model using configuration files provided in the repository. Pretrained model weights are available for download so you can run evaluations without training from scratch.

This project is primarily aimed at researchers studying autonomous driving perception or developing new 3D detection methods. It is not a finished product for deployment in a vehicle; it is a research platform for building and benchmarking detection algorithms on established datasets.