OpenPCDet Documentation

OpenPCDet supports several advanced features for 3D object detection:

• Frameworks: Both one-stage and two-stage 3D object detection frameworks.
• Distributed Computing: Distributed training and testing across multiple GPUs and multiple machines.
• Multi-scale Detection: Support for multiple heads on different scales to detect various classes.
• Point Abstraction: Stacked version set abstraction to encode varying point densities.
• Target Assignment: Adaptive Training Sample Selection (ATSS).
• Pooling Operations: RoI-aware point cloud pooling and RoI-grid point cloud pooling.
• Optimization: GPU-accelerated 3D IoU calculation and rotated NMS.

The toolbox follows specific design patterns to facilitate development and extension:

• Data-Model Separation: Uses a unified point cloud coordinate system to allow easy extension to custom datasets.
• Unified 3D Box Definition: All models use the format (x, y, z, dx, dy, dz, heading).
• Flexible Model Structure: A clear architecture that supports various 3D detection models within a single framework.

When using voxel-based detectors (e.g., SECOND, PV-RCNN, CenterPoint) or pillar-based detectors (e.g., PointPillar), ensure your configuration adheres to these constraints:

1. Z-axis: Point cloud range along z-axis / voxel_size must equal 40.
2. X & Y axes: Point cloud range along x&y-axis / voxel_size must be a multiple of 16.

You can evaluate MPPNet using a pretrained checkpoint. To avoid Out-of-Memory (OOM) errors, it is recommended to set batch_size to 1.

Single GPU Evaluation:

python test.py --cfg_file cfgs/waymo_models/mppnet_4frames.yaml --batch_size 1 --ckpt ../output/waymo_models/mppnet_4frames/default/ckpt/checkpoint_epoch_6.pth

Multi-GPU Evaluation:

bash scripts/dist_test.sh ${NUM_GPUS} --cfg_file cfgs/waymo_models/mppnet_4frames.yaml --batch_size 1 --ckpt ../output/waymo_models/mppnet_4frames/default/ckpt/checkpoint_epoch_6.pth

To visualize the predicted results during the demo, you must have either Open3D (recommended for speed) or mayavi installed. You can install them using pip:

pip install open3d
# or 
pip install mayavi

Annotations for the custom dataset must be stored in .txt files. Each line in the text file represents one bounding box in a scene using the following format:

[x y z dx dy dz heading_angle category_name]

Ensure the boxes follow the unified 3D box definition used by OpenPCDet. Example lines:

1.50 1.46 0.10 5.12 1.85 4.13 1.56 Vehicle
5.54 0.57 0.41 1.08 0.74 1.95 1.57 Pedestrian

Use test.py to evaluate models. You can specify the configuration file, batch size, and checkpoint path.

Options:

• --eval_all: Test all saved checkpoints of a specific training setting and draw performance curves on Tensorboard.
• Multi-GPU Testing: Use scripts/dist_test.sh or scripts/slurm_test_mgpu.sh.

Follow these steps to install OpenPCDet. If you have a previous version installed, you must re-install by running python setup.py develop to ensure the v0.5 components are correctly linked.

1. Clone the repository:

   git clone https://github.com/open-mmlab/OpenPCDet.git

2. Install dependencies: Install the required Python libraries listed in requirements.txt using pip.
3. Install spconv: Choose the version of spconv compatible with your PyTorch version (see requirements for details).
4. Install OpenPCDet in development mode: Run the setup script to install the library and its dependencies.

git clone https://github.com/open-mmlab/OpenPCDet.git
# (Install requirements.txt and spconv first)
python setup.py develop