CUDA Image Filters on NVIDIA TX-2 & NVIDIA GTX1070

Implementation of GPU accelerated Median Filter and Bilateral Filter for Stereo Vision using CUDA and OpenCV for CIS601 - Special Topics in Computer Architecture : GPGPU Programming

To use this repository, make sure you have the following components installed:

• OpenCV 2.4
• CUDA 8.0

Clone the repository;

Once inside the repository launch the Makefile with the "make" command;

This repository already contains some test data in the "/data" folder; If you wish to use your own images, add a link to your desired image in main.cpp located in the "/src" folder;

Performance is measured over 10 attempts, and the first warm-up kernel launch is omitted in each case. You can change this by modifying the attempts variable in main.cpp;

To run this code:

$ cd /path/to/repository/cuda-image-filters

$ make

$ ./build/filters_gpu

Results

Window Size for both filters is 9

Sigma1 for Bilateral filter is 50px

Sigma2 for Bilateral filter is 50px

Median Filter

Original Image:

CUDA Implementation:

OpenCV Implementation:

Bilateral Filter

Original Image:

CUDA Implementation:

OpenCV Implementation:

Performance

All performance times are measured in milliseconds (Y-Axis). Image resolution during performance analysis was (480 x 1200).

Median Filter

Naive Implementation

Shared Memory Implementation

Bilateral Filter

If you have any questions or find any mistakes in this repository please raise an issue or contact me.