GPU计算光流：gpu_flow

只是想迅速计算个光流图片，结果搞了大半天都没装好。

工具：
https://github.com/feichtenhofer/gpu_flow

安装步骤：

• 首先确认依赖是否都装好：
  [OpenCV 2.4] (http://opencv.org/downloads.html)
  确认版本：pkg-config --modversion opencv
  [Qt 5.4] (https://www.qt.io/qt5-4/)
  确认版本：qmake -v
  [cmake] (https://cmake.org/)

• 然后开始吧：

git clone --recursive https://github.com/feichtenhofer/gpu_flow
cd gpu_flow
mkdir build
cd build
cmake -DCUDA_USE_STATIC_CUDA_RUNTIME=OFF ..
make

problems

• cannot find lopencv_dep_cudart

https://github.com/opencv/opencv/issues/6542

在CMakeLists.txt中：FIND_PACKAGE(OpenCV REQUIRED )之前添加:

set(CUDA_USE_STATIC_CUDA_RUNTIME OFF)

参考链接：
http://www.cnblogs.com/darkknightzh/p/5638117.html
http://blog.csdn.net/qq_36130482/article/details/70171301
http://blog.csdn.net/u012526003/article/details/64500596

修改后的CMakeLists.txt:

# OpenCV Config
set(CUDA_USE_STATIC_CUDA_RUNTIME OFF)
find_package( OpenCV REQUIRED )
message(STATUS " version: ${OpenCV_VERSION}")
message("-- OPENCV include: " ${OpenCV_INCLUDE_DIRS})
message("-- OPENCV libs dir: " ${OpenCV_LIB_DIR})
message("-- OPENCV libs: " ${OpenCV_LIBS} )

试了几次还是不行，于是

rm build -rf
mkdir build && cd build
cmake ..
make

还是不行。
于是：https://github.com/caffe2/caffe2/issues/303

rm -rf build
mkdir build && cd build
cmake -DCUDA_USE_STATIC_CUDA_RUNTIME=OFF ..
make

可以了！感人。

• ``lib/libstdc++.so.6: version GLIBCXX_3.4.20 not found (required by /usr/local/lib/libopencv_gpu.so.2.4)

http://blog.csdn.net/xiaolong2w/article/details/23915171
我的libstdc++.so.6是anaconda3里面的，用strings libstdc++.so.6查看了一下（忘了存结果）：
只到3.4.19。
所以我升级了一下我的conda和所有的库：

conda update conda  #更新conda本身
conda update --all  #更新所有库，有点慢

现在：

root@Amax-02:~/anaconda3$ strings /home/root/anaconda3/lib/libstdc++.so.6 | grep GLIBCXX
GLIBCXX_DEBUG_MESSAGE_LENGTH
GLIBCXX_3.4
GLIBCXX_3.4.1
GLIBCXX_3.4.2
GLIBCXX_3.4.3
GLIBCXX_3.4.4
GLIBCXX_3.4.5
GLIBCXX_3.4.6
GLIBCXX_3.4.7
GLIBCXX_3.4.8
GLIBCXX_3.4.9
GLIBCXX_3.4.10
GLIBCXX_3.4.11
GLIBCXX_3.4.12
GLIBCXX_3.4.13
GLIBCXX_3.4.14
GLIBCXX_3.4.15
GLIBCXX_3.4.16
GLIBCXX_3.4.17
GLIBCXX_3.4.18
GLIBCXX_3.4.19
GLIBCXX_3.4.20
GLIBCXX_3.4.21

可以啦~

其他的在作者的README.md都写的很清楚啦：

Configuration:

You should adjust the input and output directories by editing the variables vid_path, out_path and out_path_jpeg in compute_flow.cpp. Note that these folders have to exist before executing.

Usage:

./brox_flow [OPTION]...

Available options:
* start_video: start with video number in vid_path directory structure [1]
* gpuID: use this GPU ID [0]
* type: use this flow method Brox = 0, TVL1 = 1 [1]
* skip: the number of frames that are skipped between flow calcuation [1]

Additional features in compute_flow.cpp:
* float MIN_SZ = 256: defines the smallest side of the frame for optical flow computation
* float OUT_SZ = 256: defines the smallest side of the frame for saving as .jpeg
* bool clipFlow = true;: defines whether to clip the optical flow larger than [-20 20] pixels and maps the interval [-20 20] to [0 255] in grayscale image space. If no clipping is performed the mapping to the image space is achieved by finding the frame-wise minimum and maximum displacement and mapping to [0 255] via an adaptive scaling, where the scale factors are saved as a binary file to out_path.

Example:

./brox_flow gpuID=0 type=1