做个简单测试：
In [16]: def calc():
    ...:     frame = np.random.randint(0, 256, (2160, 3840, 3)).astype(np.uint8)
    ...:     kernel = np.ones((31,31), np.uint8)
    ...:     cv2.erode(frame, kernel, iterations=10)
    ...:

In [17]: %timeit calc()
195 ms ± 1.69 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

然后16线程执行下：
In [18]: def multi_thread():
   ...:     thres = [threading.Thread(target=calc) for _ in range(16)]
   ...:     [x.start() for x in thres]
   ...:     [x.join() for x in thres]
   ...:

In [19]: %timeit multi_thread()
1.37 s ± 18.3 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

同时跑的时候看了下cpu占用率是240%左右。
有点奇怪。加速了，但这个加速比不太对。

再测试下多进程的情况：
In [21]: def multi_proc():
    ...:     thres = [multiprocessing.Process(target=calc) for _ in range(16)]
    ...:     [x.start() for x in thres]
    ...:     [x.join() for x in thres]
    ...:

In [22]: %timeit multi_proc()
614 ms ± 20.1 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

好多了但还是有3倍差距，我这是16核cpu，感觉差距不应该有这么大...

anyway，opencv release gil的行为看来没有我想的那么乐观。

【 在 CKevin 的大作中提到: 】
: 你是说，比如
: def calc(image):
:   # cv2 and np here
: ...................
[upload=1][/upload]
--
修改:lvsoft FROM 180.158.52.*
FROM 180.158.52.*