python - 为什么不能发送cv2视频源？

Question

我尝试使用 MPI 分散将 cv2 视频捕获源发送到不同的进程。这样，每个进程都会有一个句柄来检索帧。

if rank == 0:
  video_source = cv2.VideoCapture('my_video.avi')
  source_list = [video_source] * world_communicator.Get_size()
else:
  source_list = None
data = world_communicator.scatter(source_list, root=0)
# Do stuff with access to the video source...

但是，此代码出现以下错误。

[dfaux@ubiquitous mpi_practice]$ mpirun -n 2 python cv2_video_practice.py
Traceback (most recent call last):
  File "cv2_video_practice.py", line 93, in <module>
    process_video(source, destination, kernel)
  File "cv2_video_practice.py", line 58, in process_video
    data = comm.scatter(frame_counts, root=0)
  File "Comm.pyx", line 874, in mpi4py.MPI.Comm.scatter (src/mpi4py.MPI.c:68023)
  File "pickled.pxi", line 656, in mpi4py.MPI.PyMPI_scatter (src/mpi4py.MPI.c:32402)
  File "pickled.pxi", line 131, in mpi4py.MPI._p_Pickle.dumpv (src/mpi4py.MPI.c:26847)
  File "pickled.pxi", line 81, in mpi4py.MPI._p_Pickle.dump (src/mpi4py.MPI.c:26188)
cPickle.PicklingError: Can't pickle <type 'cv2.VideoCapture'>: it's not the same object as cv2.VideoCapture

为什么我不能 pickle 视频源？为什么是<type 'cv2.VideoCapture'>与 cv2.VideoCapture 不是同一个对象？

Answer 1

cv2.VideoCapture 不是所有向量的列表。相反，它是一个指向视频中特定帧的句柄。这样，将这样的对象分散到一堆其他进程中就没有意义了。相反，我建议在每个进程中打开一个句柄，然后让每个进程从不同的框架开始:

source.set(cv.CV_CAP_PROP_POS_FRAMES, start_frame)