Faster JPEG encode/decode processing

OpenCV default JPEG compression/decompression is way too slow, we need one faster implement

restart_interval matters as GPU parallel processing relaying on it, CESNET/GPUJPEG/issues/29. Please tweak it for specific image size. (8, 16, 32) can be used for 1280x720.

Benchmark

Time with 1280x720 images and restart_interval=16 on Xavier and Allienware([email protected]×8, GeForce GTX 1080 with Max-Q). Note, the speed varies for different running, however, the speedup ratio is still satisfying.

./demo_benchmark.sh

Time ms (100 iterations)

C++ version:

Method	Xavier	GTX1080(Max-Q)
OpenCV Encode	2369	739
GPUJPEG Encode	245	152
OpenCV Decode	1884	829
GPUJPEG Decode	429	206

Python version:

Method	Xavier	GTX1080(Max-Q)
OpenCV Encode	791	704
GPUJPEG Encode	242	154
OpenCV Decode	1269	1092
GPUJPEG Decode	492	263

Usage

# Build and install GPUJPEG first (CMake)
git clone https://github.com/CESNET/GPUJPEG.git
...

# Build this demo (CMake)
# Please change the OpenCV (and libgpujpeg) path
git clone https://github.com/MiaoDX/GPUJPEG_demo.git --depth 1
git submodule update --init --depth 1
...

./demo_benchmark.sh

Optional choices

• CESNET/GPUJPEG
  • Current choice, open-source, quick enough
• Fastvideo
  • Close-sourced SDK, many other functions
  • Watermark with the trial version
• nvJPEG
  • Not available for Xavier currently (20191016)

Changelog

• 2019.11.28, decode, python binding, and benchmark
• 2019.11.27, refactoring and python binding for encode with pybind
• 2019.10.16, initial version, example codes for encode

TODO

• Use OpenCV cv::cuda::GpuMat directly