A PyTorch toolbox that corrals point clouds into detectable objects

What it does Det3D is a PyTorch-based toolbox that bundles implementations of voxel and point-based 3D object detectors—SECOND, PointPillars, and VoxelNet—under one roof. It targets the standard autonomous-driving benchmarks (KITTI, nuScenes, and Lyft) and ships with configuration files and checkpoint links that reproduce published numbers, including the CBGS nuScenes result. The codebase also layers on distributed-training utilities, synchronized batch normalization, and a built-in visualizer so you are not left plotting .bin files by hand.

The interesting bit The project essentially tries to do for 3D detection what mmdetection did for 2D, openly acknowledging its debt to that ecosystem and to second.pytorch. Its value is in the plumbing: multi-dataset loaders, flexible anchor dimensions, and rotated RoI align bundled with training hooks that actually work across nodes.

Key highlights

• Model zoo with reported KITTI and nuScenes metrics (CBGS hits 61.3 NDS; SECOND posts 87.96 3D AP on KITTI car).
• Supports both point-based and voxel-based architectures in the same framework.
• Distributed data parallel and SyncBN built in.
• Self-contained visualization and TensorboardX logging.
• Multi-scale training and finetuning hooks.

Caveats

• Several promised models—PointRCNN and PIXOR—remain unchecked on the TODO list and are not actually available.
• Waymo is listed as a dataset but the section is empty, and the maintainers explicitly ask for contributions to support it.
• The README contains duplicate entries and typos (“licenes”, “INSTALATION”), which hints at sporadic maintenance.

Verdict Worth a look if you need a unified PyTorch starting point for KITTI or nuScenes 3D detection and do not mind filling in a few gaps yourself. Skip it if you need the still-unreleased models or broad dataset coverage beyond what is already implemented.