PointPillars-ROS

A 3D detection Pointpillars ROS deployment on Nvidia Jetson TX1/TX2

This repo implements https://github.com/hova88/PointPillars_MultiHead_40FPS into Autoware lidar_point_pillars framework https://github.com/autowarefoundation/autoware_ai_perception/tree/master/lidar_point_pillars.

However, multihead 40FPS models is originally tested on 3080Ti. It takes about 700ms one frame on Nvidia TX1, while 140ms on Nvidia Xavier.

I use OpenPCDet to train accelerated models within 250ms on TX1 (model: zz0808_256_e50).

Features

It supports 11/10 gather-feature point-pillar models. 11 gfeature model is trained by Nuscenes Data by OpenPCDet, which has 5 basic features. 10 gfeature model is trained by Kitti Data by OpenPCDet, which has 4 basic features.

However, the INPUT feature numbers of both dataset are 5. So, different models can deal with different input from rosbags.

Update

The post process of the original PointPillars_MultiHead_40FPS project, is HARD CODED some configuraion.

Hard coded 10 heads in 6 groups, I change it to read from yaml.
Hard coded feature_num of the pillar with 5, I change it to read from yaml (Kitti = 4, Nuscence = 5).
Hard coded gather_feature_num with 11, I change it to feature_num + 6.

If you use kitti dataset to train a 11 gfeature model (you can add a refill zero dim into training procedure. While, I upload one sample model), you can use pointpillar_kitti_g11.yaml to infer this model. The differences of the yaml file is: WITH_REFILL_DIM: True

Requirements

My Environment

Ubuntu 18.04

ROS Melodic

TX1

jetson_release
 - NVIDIA Jetson TX1
   * Jetpack 4.5.1 [L4T 32.5.1]
   * NV Power Mode: MAXN - Type: 0
   * jetson_stats.service: active
 - Libraries:
   * CUDA: 10.2.89
   * cuDNN: 8.0.0.180
   * TensorRT: 7.1.3.0
   * Visionworks: 1.6.0.501
   * OpenCV: 3.4.5 compiled CUDA: YES
   * VPI: ii libnvvpi1 1.0.15 arm64 NVIDIA Vision Programming Interface library
   * Vulkan: 1.2.70

Xavier

 - NVIDIA Jetson AGX Xavier [16GB]
   * Jetpack 4.5.1 [L4T 32.5.1]
   * NV Power Mode: MAXN - Type: 0
   * jetson_stats.service: active
 - Libraries:
   * CUDA: 10.2.89
   * cuDNN: 8.0.0.180
   * TensorRT: 7.1.3.0
   * Visionworks: 1.6.0.501
   * OpenCV: 3.4.5 compiled CUDA: YES
   * VPI: ii libnvvpi1 1.0.12 arm64 NVIDIA Vision Programming Interface library
   * Vulkan: 1.2.70

dependence

Could not find the required component 'jsk_recognition_msgs'.

sudo apt-get install ros-melodic-jsk-recognition-msgs 
sudo apt-get install ros-melodic-jsk-rviz-plugins

Need OpenCV compiled with CUDA.

Usage

How to compile

Simply, use catkin_make build up the whole project.

Add project to runtime environment.

source devel/setup.bash

How to launch

Launch file (cuDNN and TensorRT support):

pfe_onnx_file, rpn_onnx_file, pp_config, input_topic are required

roslaunch lidar_point_pillars lidar_point_pillars.launch pfe_onnx_file:=/PATH/TO/FILE.onnx rpn_onnx_file:=/PATH/TO/FILE.onnx pp_config:=/PATH/TO/pp_multihead.yaml input_topic:=/points_raw

score_threshold, nms_overlap_threshold, etc are optional to change the runtime parameters.

Or, simply,

Use launch.sh to run.

Test launch

roslaunch test_point_pillars test_point_pillars.launch

nuscenes test data download: nuscenes_10sweeps_points.txt

From: https://github.com/hova88/PointPillars_MultiHead_40FPS

Tx1 (single test):

  Preprocess    7.48567  ms
  Pfe           266.516  ms
  Scatter       1.78591  ms
  Backbone      369.692  ms
  Postprocess   35.8309  ms
  Summary       681.325  ms

Xavier (single test):

  Preprocess    2.15862  ms
  Pfe           46.2593  ms
  Scatter       0.54678  ms
  Backbone      80.7096  ms
  Postprocess   11.3462  ms
  Summary       141.034  ms

Test Rosbag:

I use nuscenes2bag to create some test rosbag: nu0061 all 19s 5.5G, download password: s2eh, nu0061 laser and tf only 19s 209M, download password: m7wh.

To use this nuscenes rosbag, you shoulde change input_topic to /lidar_top , and use src/rviz/nuscenes.rviz for visualization.

Usually, I use rosbag play r 0.1 for more play time.

More test rosbag, like kitti, carla or real data by myself, will be released recently.

Models Files:

Faster ONNX models on TX1:

zz0809_512_e50 model is with the same config file as cbgs model, and the evaluation data is re-tested by the same eval benchmark.
zz0808_256_e50 model is half resolution, you should used this config file to run: src/lidar_point_pillars/cfgs/tx1_ppmh_256x256.yaml
z0927_kitti is trained by kitti dataset, with three classes. It has only 10 (4+6) gather features, and can run with this config file: src/lidar_point_pillars/cfgs/pointpillar_kitti_g10.yaml
z1009_kitti_g11 is trained by kitti dataset, with three classes. It has 11 gather features, with one refile zero dim. It can run with this config file: src/lidar_point_pillars/cfgs/pointpillar_kitti_g11.yaml

	download	Tx1 time	Xavier time	resolution	training data	mean ap	nd score	car ap	ped ap	truck ap
cbgs_ppmh	pfe backbone	~700ms	~140ms	64x512x512	unknown	0.447	0.515	0.813	0.724	0.500
zz0809_512_e50	pfe backbone	~700ms	~140ms	64x512x512	nusc tr-v	0.460	0.524	0.818	0.733	0.507
zz0808_256_e50	pfe backbone	~250ms	~110ms	64x256x256	nusc tr-v	0.351	0.454	0.781	0.571	0.427
kitti models								car ap@0.7	ped ap@0.5	truck ap@0.7
z0927_kitti_g10	pfe backbone	~700ms	~140ms	64x512x512	kitti			90.149	44.893	34.977
z1009_kitti_g11_e72	pfe backbone	~700ms	~140ms	64x512x512	kitti			90.191	46.915	40.944

More models will be released recently.

mrhagchwh / PointPillars-ROS