Point Pillars for 3D Object Detection: ver. 1.0

Autoware package for Point Pillars. Referenced paper.

This node can be compiled either using cuDNN and TensorRT or by using TVM, the default are cuDNN and TensorRT.

Requirements

• CUDA Toolkit v9.0 or v10.0

To compile the node using the cuDNN and TensorRT support the requirements are:

• cuDNN: Tested with v7.3.1

• TensorRT: Tested with 5.0.2 -> How to install

To compile the node using TVM support the requirements are:

• TVM runtime, TVM Python bindings and dlpack headers

• tvm_utility package

How to setup

Setup the node using cuDNN and TensorRT support:

1. Download the pretrained file from here.

$ git clone https://github.com/k0suke-murakami/kitti_pretrained_point_pillars.git

Setup the node using TVM support:

1. Clone the modelzoo repository for Autoware here.

2. Use the TVM-CLI to export point pillars models to TVM (Instruction on how to do it are present in the repository

   ```), models are 
   ```perception/lidar_obstacle_detection/point_pillars_pfe/onnx_fp32_kitti
   ``` and 
   ```perception/lidar_obstacle_detection/point_pillars_rpn/onnx_fp32_kitti
   

3. Copy the generated files into the tvm_models/tvm_point_pillars_pfe and tvm_models/tvm_point_pillars_rpn folders respectively for each model. With these files in place, the package will be built using TVM.

4. Compile the node.

How to launch

• Launch file (cuDNN and TensorRT support): roslaunch lidar_point_pillars lidar_point_pillars.launch pfe_onnx_file:=/PATH/TO/FILE.onnx rpn_onnx_file:=/PATH/TO/FILE.onnx input_topic:=/points_raw

• Launch file (TVM support): roslaunch lidar_point_pillars lidar_point_pillars.launch

• You can launch it through the runtime manager in Computing tab, as well.

API

/**
* @brief Call PointPillars for the inference.
* @param[in] in_points_array pointcloud array
* @param[in] in_num_points Number of points
* @param[out] out_detections Output bounding box from the network
* @details This is an interface for the algorithm.
*/
void doInference(float* in_points_array, int in_num_points, std::vector<float> out_detections);

Parameters

Outputs

Notes

• To display the results in Rviz objects_visualizer is required. (Launch file launches automatically this node).

• Pretrained models are available here, trained with the help of the KITTI dataset. For this reason, these are not suitable for commercial purposes. Derivative works are bound to the BY-NC-SA 3.0 License. (https://creativecommons.org/licenses/by-nc-sa/3.0/)

Changelog for package lidar_point_pillars

1.11.0 (2019-03-21)

• [Feature]PointPillars (#2029)
• Contributors: Kosuke Murakami