Why is `capture.set(cv2.CAP_PROP_POS_FRAMES, i)` so slow?

Python
1

I am trying to sample every 5th frame from a source. I can either iterate through the frames and pick every 5th frame and discard the rest like:

def sample_video_loop(capture, sampling_gap=5, batch_size=5):
    frames = []
    for i in range(batch_size):
        for j in range(sampling_gap):
            ret, f0 = capture.read()
        frames += [f0]
    return frames

Or I can use the VideoCapture set() method to set the property cv2.CAP_PROP_POS_FRAMES to like:

def sample_video_set_pos(capture, sampling_gap=5, batch_size=5):
    frames = []
    pos = 0
    for i in range(batch_size):
        ret, f0 = capture.read()
        pos += sampling_gap
        capture.set(cv2.CAP_PROP_POS_FRAMES, pos)
        frames += [f0]
    return frames

Without checking the implementation ideally the latter method should be faster. But when i timed it, it is 3-4 times slower.

%%timeit
frames = sample_video_loop(dataset, 5, 5)

>>> 98.4 ms ± 5.49 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

%%timeit
frames = sample_video_set_pos(dataset, 5, 5)

>>> 305 ms ± 16.5 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

Is this a bug or am i missing something? If not then what is the best practice to sample frames?

Thanks in advance.

2

no bug here, “random access” in video IS slow.
whenever you send it to frame x, it has to decode from the previous to the next key frame, to deliver a single image.

i’d say, you found out already :wink:

1 Like
3

That’s a bit disappointing but makes sense. Thanks.