如何在python中使用cv2和多处理并行获取视频帧

Question

I've been working with cv2 & multiprocessing in python, and I finally have a working script that does stuff to the individual frames once they are already in an input queue. However, I wanted to speed up getting the frames into the queue in the first place by using multiple cores, so I tried to use the same multiprocessing approach to read each image into the queue. I can't seem to get this to work though, and I'm not sure why. I thought maybe it was because I was trying to write to one queue, so I split those up, but now I'm wondering if it's because I'm trying to read from the same video file at the same time.

Here is what I am hoping to accomplish in pseudocode:

for process in range(processCount):
    start a process that does this:
        for frame in range(startFrame,endFrame):
            set next frame to startFrame
            read frame
            add frame to queue

Here is my current code. I've tried using pool & separate processes, but for now I'm sticking to separate processes because I'm not sure if the problem is with queue management. If I call getFrame manually, I get the right stuff into the queue, so I think that function by itself works okay.

I'm sure I'm doing something really silly (or really odd). Can someone suggest a solution? It would be great to just have one queue as well... I just had two to try to break down the problem.

Thanks in advance.

import numpy as np
import cv2
import multiprocessing as mp
import time

def getFrame(queue, startFrame, endFrame):
    for frame in range(startFrame, endFrame):
        cap.set(1,frame)
        frameNo = int(cap.get(0))
        ret, frame = cap.read()
        queue.put((frameNo,frame))

file = 'video.mov'
cap = cv2.VideoCapture(file)
fileLen = int(cap.get(7))

# get cpuCount for processCount
processCount = mp.cpu_count()/3

inQ1 = mp.JoinableQueue()  # not sure if this is right queue type, but I also tried mp.Queue()
inQ2 = mp.JoinableQueue()
qList = [inQ1,inQ2]

# set up bunches
bunches = []
for startFrame in range(0,fileLen,fileLen/processCount):
    endFrame = startFrame + fileLen/processCount
    bunches.append((startFrame,endFrame))

getFrames = []
for i in range(processCount):
    getFrames.append(mp.Process(target=getFrame, args=(qList[i], bunches[i][0],bunches[i][1],)))

for process in getFrames:
    process.start()

results1 = [inQ1.get() for p in range(bunches[0][0],bunches[0][1])]
results2 = [inQ2.get() for p in range(bunches[1][0],bunches[1][1])]

inQ1.close()
inQ2.close()
cap.release()

for process in getFrames:
    process.terminate()
    process.join()

Answer 1

There is indeed a mistake in the code : the use of the same VideoCapture object across processes. Obviously there's a conflict on the position currently being read in the file.

This being said, when trying to instantiate one VideoCapture per process, my interpreter crashes (tested with python3.4.2 + opencv3.0.0-beta , and python2.7.6 + opencv2.4.8 ). Here's my try so far if you want to check it / go further.

import cv2
import multiprocessing as mp

def getFrame(queue, startFrame, endFrame):
    cap = cv2.VideoCapture(file)  # crashes here
    print("opened capture {}".format(mp.current_process()))
    for frame in range(startFrame, endFrame):
        # cap.set(cv2.CAP_PROP_POS_FRAMES, frame)  # opencv3            
        cap.set(cv2.cv.CV_CAP_PROP_POS_FRAMES, frame)
        # frameNo = int(cap.get(cv2.CAP_PROP_POS_FRAMES))  # opencv3
        frameNo = int(cap.get(cv2.cv.CV_CAP_PROP_POS_FRAMES))
        ret, f = cap.read()
        if ret:
            print("{} - put ({})".format(mp.current_process(), frameNo))
            queue.put((frameNo, f))
    cap.release()

file = "video.mov"
capture_temp = cv2.VideoCapture(file)
# fileLen = int((capture_temp).get(cv2.CAP_PROP_FRAME_COUNT))  # opencv3
fileLen = int((capture_temp).get(cv2.cv.CV_CAP_PROP_FRAME_COUNT))
capture_temp.release()

# get cpuCount for processCount
# processCount = mp.cpu_count() / 3
processCount = 2

inQ1 = mp.JoinableQueue()  # not sure if this is right queue type, but I also tried mp.Queue()
inQ2 = mp.JoinableQueue()
qList = [inQ1, inQ2]

# set up bunches
bunches = []
for startFrame in range(0, fileLen, int(fileLen / processCount)):
    endFrame = startFrame + int(fileLen / processCount)
    bunches.append((startFrame, endFrame))

getFrames = []
for i in range(processCount):
    getFrames.append(mp.Process(target=getFrame, args=(qList[i], bunches[i][0], bunches[i][1])))

for process in getFrames:
    process.start()

results1 = [inQ1.get() for p in range(bunches[0][0], bunches[0][1])]
results2 = [inQ2.get() for p in range(bunches[1][0], bunches[1][1])]

inQ1.close()
inQ2.close()

for process in getFrames:
    process.terminate()
    process.join()