D435i: How to generate PCL Point Cloud at faster rate?

[KainanSu]

Like the issue that's been closed. #7676

I want to process pointcloud with pcl, but the example in this repo convert data format to pcl costs lots time. (600ms)
In #7676, that is say rs-gl may use gpu help to accelerate converting.

But I meet 2 problem as follow:

1. I cannot find anything help to convert with gpu, but just use gpu to store pointcloud(rs2::gl::pointcloud). I dont think it can help me to accelerate.
2. Latest SDK doesnt include librealsense2-gl/rs_processing_gl.hpp.

Could you give any solutions? Thanks.

[MartyG-RealSense]

Hi @KainanSu Pointclouds can be a processing intensive operation, so reducing the burden on the CPU (lowering its 'percentage usage) by offloading the processing work onto a GPU can in itself improve an application's performance.

GLSL can work with RS2:: instructions simply by inserting GL:: into the instruction. For example for pointclouds, rs2::pointcloud becomes rs2::gl::pointcloud

An example of rs2::pointcloud in a RealSense PCL pointcloud script is rs-pcl-color

https://github.com/IntelRealSense/librealsense/blob/master/wrappers/pcl/pcl-color/rs-pcl-color.cpp#L158

However, if the computer is a low-power device that does not have a strong GPU then using GLSL may not provide a noticeable performance boost.

More information about GLSL is at #3654

The file that you are searching for is in the librealsense > include > librealsense2_gl folder of the SDK in the latest master version.

https://github.com/IntelRealSense/librealsense/tree/master/include/librealsense2-gl

[KainanSu]

@MartyG-RealSense Hi, Marty! Thanks for such a quick response!

I have reinstall the SDK twice, but I cannot see librealsense2_gl in C:\Program Files (x86)\Intel RealSense SDK 2.0\include. There is just librealsense2 In the foder.

When I install SDK, I had select full install. Buy the way , I use win10, and CUDA 9.0.

[MartyG-RealSense]

If you installed the SDK on Windows with the Assets file list's Intel.RealSense.SDK-WIN10 installer file, there is another link in the list called Source code - zip. If that zip file is downloaded and extracted then there should be a similar folder and file structure to the GitHub website version of the SDK that I linked to.

The SDK can be built from source code on Windows using instructions at the link below, but as far as I am aware GLSL support should be in the SDK version installed by the Intel.RealSense.SDK-WIN10 installer file too.

https://github.com/IntelRealSense/librealsense/blob/master/doc/installation_windows.md

CUDA support requires a computer with an Nvidia GPU, whilst GLSL is 'vendor neutral' and should work with any GPU brand.

If you have a Windows PC with an Nvidia video card then I would think that CUDA support could be enabled by building the SDK from source code with CMake using the instructions link above and setting the CMake build flag BUILD_WITH_CUDA to true. I'm not certain though, as virtually all librealsense cases that use CUDA are on computers running Linux, mostly on Nvidia Jetson computing boards that have a built-in Nvidia GPU.

[KainanSu]

@MartyG-RealSense Thanks, Marty. I try it right now.

[KainanSu]

@MartyG-RealSense Hi, Marty! I looked into the convert code and I found the function RGB_Texture costs most time(80%).
I try to rewrite RGB_Texture and now it can capture frame in 200ms, it is already enough for me.

[MartyG-RealSense]

That's great to hear that you achieved a fast-working solution, @KainanSu - thanks very much for the update!

[MartyG-RealSense]

Case closed due to solution achieved and no furtther comments received.

[FeiSong123]

Hello, I heard that you have rewritten RGB_Texture function then improves the speed of point cloud acquisition. Can you share your rewritten function, please？ @KainanSu

[KainanSu]

@FeiSong123 Sorry, I cannot find the code. But I remember the code I rewrite were below: reduce function calls, reduce object unnecessary recreations.

[FeiSong123]

All right. Thank you for your help. I'll go back and try your way.

[FeiSong123]

Hello, I tried the method you said, and then successfully reduced the time of point cloud acquisition. Thank you for your help! @KainanSu