Realsense disparity refinement with CUDA

I made the disparity refinement project for Intel Realsense D400 series. It's based on the C++ & CUDA. I did the disparity re-computation about the IR pair(left & right) from Intel Realsense D435.

I referred "SOS: Stereo Matching in O(1) with Slanted Support Windows" and "HITNet: Hierarchical Iterative Tile Refinement Network for Real-time Stereo Matching" which are based on the tile disparity.

• SOS paper : https://ieeexplore.ieee.org/abstract/document/8593800
• HITNet paper : https://arxiv.org/abs/2007.12140

The environment for testing is below.

• Intel Realsense D435 camera ([848 x 480], [1280 x 720])
• OS : Windows 10
• IDE : Visual studio 2015 community
• CPU : Intel(R) Core(TM) i7-9700K (3.60GHz)
• GPU : Geforce RTX 2080 ti
• RAM : 64 GB

Dependency for testing

• Opencv 4.1.0
• Realsense SDK 2.24
• SDL2 : https://www.libsdl.org/download-2.0.php
• glew : http://glew.sourceforge.net/
• CUDA 10.1

The method consists of several modules

• (1) Census transform.
• (2) Disparity to tile.
• (3) Tile disparity refinement using the parabola fitting.
• (4) Tile slant estimation using the eigen value decomposition.
• (5) per pixel estimation based on the tile.

Tile slant visualization

Disparity refinement before/after

• Before

• After

TODO

Please use this method and give me a comment about this.

• I'm trying to change the build type to cmake.