RidgeRun OpenCV Fork

Introduction

The RidgeRun OpenCV fork is a modified version of the project with various improvements around speed and efficiency. While some changes are general (i.e.: any platform will benefit from them), others are specific to the NVIDIA Jetson family.

The fork is hosted at:

https://github.com/RidgeRun/opencv

Building the Project

Follow the instructions in our Compiling OpenCV from Source page. Make sure you select:

• RidgeRun fork
• GStreamer support
• CUDA support (if applicable)

Enhancements

GStreamer Video Capture

 STATUS: Submitted.

 BRANCH: zerocopy-gstreamer-videocapture

Effectively removes the memory copy when transferring data from the GstBuffer to the cv::Mat. Now the GstBuffer and its associated memory will remain alive throughout the lifespan of the matrix.

Tested under the following conditions

• OpenCV Version: 4.4.0 and 4.5.2
• FPS and CPU usage taken with the GstPerf element.
• Jetson Nano and x86 PC
• The following source:

#include <opencv2/opencv.hpp>

int main() {
  cv::VideoCapture cap(
     "fakesrc sizetype=fixed filltype=zero sizemax=24883200 ! video/x-raw,format=BGR,width=3840,height=2160 ! "
     "perf print-arm-load=true ! appsink drop=false sync=false max-buffers=3", 
     cv::CAP_GSTREAMER);
  
  while(1){
    cv::Mat frame;
    cap >> frame;	
  }

  cap.release();
  return 0;
}

• Run with

g++ -o benchmark benchmark.cc $(pkg-config --cflags --libs opencv4) -std=c++11
GST_DEBUG=perf:4 ./benchmark

OpenCV 4.4.0 Results on Jetson Nano

GStreamer Video Capture - Max Framerate

OpenCV 4.5.2 Results on x86 PC

GStreamer Video Capture - Max Framerate

GStreamer Main Loop

 STATUS: Accepted!

 BRANCH: optional-gstreamer-mainloop

Optionally enables a GMainLoop when capturing using GStreamer Video Capture.

A running GMainLoop processes many events on the GLib/GStreamer world. While some things may work without it, many others won't. Examples of these are signals, timers and many other source events. The problem becomes more concerning by the fact that some GStreamer elements rely on signals to work. This branch allows the user to specify an OpenCV option to start a main loop, if needed. Since the loop blocks, this is done in a separate thread.

Use the following environment variable to enable the main loop:

 OPENCV_VIDEOIO_GSTREAMER_START_MAINLOOP=true

Tested under the following conditions

• OpenCV Version: 4.4.0
• Mainloop used by GstShark tracers.
• Jetson Nano
• The following source:

#include <opencv2/opencv.hpp>

int main() {
  cv::VideoCapture cap("videotestsrc is-live=true ! appsink", cv::CAP_GSTREAMER);
  
  while(1){
    cv::Mat frame;
    cap >> frame;	
  }

  cap.release();
  return 0;
}

• Run with

g++ -o mainloop mainloop.cc $(pkg-config --cflags --libs opencv4) -std=c++11
OPENCV_VIDEOIO_GSTREAMER_START_MAINLOOP=true GST_DEBUG=GST_TRACER:7 GST_TRACERS=framerate ./mainloop

Note how, if the mainloop is not enabled, the tracer won't output any data.