yucedagonurcan / ImageCudaPrepRosNode

This is an implementation for preprocessing steps done for image data on CUDA using NPP.

CUDA Image Processing ROS Node

This is a ROS implementation of a fairly common image processing pipeline with using CUDA and NPP:

1. Debayer
2. Undistort(Rectify)
3. Resize

Used Packages & Resources

1. NPP library for Debayer & Resize classes
2. Code structure and function of Undistort class are highly related to the dusty-nv/jetson-utils

Run in your machine

1. Go to nearest ROS workspace and clone this repo:

cd catkin_ws/src
git clone https://github.com/yucedagonurcan/ImageCudaPrepRosNode.git

2. Build the node:

catkin build cuda_img_processing --cmake-args -DCMAKE_BUILD_TYPE=Release

Benchmarks & Thoughts

• I don't think this pipeline alone can make a drastic difference in terms of throughput and data shows that too:
  • Delay per image (CPU) = 0.055
  • Delay per image (GPU) : 0.050
• The most important metric in this project can be freeing the CPU from this pipeline:
  • CPU usage ( CPU ): 26.5%
  • CPU usage ( GPU ): 11.5%
  • I am pretty sure that we can further optimize this code too.

Example run