Home

Awesome

SE-SDD-AI-TRT

SE-SSD-AI-TRT(SE-SSD ALL IN TensorRT,NMS not implemented in TensorRT,implemented in c++)

SE-SSD consists of six parts:

Config

How to Run

  1. build SE-SSD-AI-TRT and run
firstly, install TensorRT,my environment is ubuntu 18.04, cuda 10.2,cudnn8.2.
I installed TensorRT with TensorRT-8.0.1.6.Linux.x86_64-gnu.cuda-10.2.cudnn8.2.tar.gz.

after that, modify CMakeLists.txt
include_directories(/home/xxx/softwares/nvidia/TensorRT-8.0.1.6/include)
link_directories(/home/xxx/softwares/nvidia/TensorRT-8.0.1.6/lib)
Change these two lines to your own path

cd SE-SSD-AI-TRT
mkdir build
cd build
cmake ..
make
sudo ./se-ssd-ai-trt -s             // serialize model to plan file i.e. 'se-ssd-ai-trt.engine'
sudo ./se-ssd-ai-trt -d    // deserialize plan file and run inference, lidar points will be processed.
predicted outputs saved in SE-SSD-AI-TRT/data/outputs folder

one frame takes about 1-2 seconds on my laptop with Intel(R) Core(TM) i5-7300HQ and NVIDIA GeForce GTX 1050 Mobile(1050ti), it is very slow, needs to be optimized in the future.

  1. show predicted 3D boxes in the lidar frame
fristly install anaconda.
then,
1 conda create -n pc_show python=3.6
2 conda activate pc_show
3 pip install vtk==8.1.2
4 pip install mayavi
5 pip install PyQt5
6 pip install opencv-python

cd tools
python show_box_in_points.py

if An error occurred like: qt.qpa.plugin: Could not load the Qt platform plugin "xcb" in "/home/xxx/anaconda3/envs/pc_show/lib/python3.6/site-packages/cv2/qt/plugins" even though it was found.
just delete it by run: sudo rm -rf /home/xxx/anaconda3/envs/pc_show/lib/python3.6/site-packages/cv2/qt/plugins
and try again

warning: do not close current Mayavi Scene window, type c in running terminal and press Enter,
it will show next lidar frame with predited 3d boxes in current Mayavi Scene window. 

Image text Image text

More Information

Reference code:

SE-SSD

spconv

tensorrtx

tensorrt_plugin

CUDA-PointPillars

frustum_pointnets_pytorch