This tells me a lot of the overhead is just in ffmpeg initializing, opening the file, etc.Ī very naive approach might be to extract all frames to disk and then pick off the ones desired. Therefore, if I wanted say 40 frames from this file, I'd be waiting at most 12 seconds, whereas the entire file's decode time is half that.

The same file can be decoded to null in only 6 seconds. For only three frames it might not be too bad, but if I wanted to extract a list of, say, 100 or 200 or more frames, it quickly gets unwieldy even for files with good keyframes.Ī very rudimentary test showed that for a test file, five minutes long, with keyframes every 5 seconds, each invocation of ffmpeg takes an average of about 0.22 seconds to run, with the fastest being 0.11 and the slowest being 0.34 seconds. If you assume the worst case (no usable keyframes) then the most efficient way to extract the frames would be to decode the stream once and pick off each frame as it occurs. The problem with this approach is it seems to be resource intensive for a few reasons: running ffmpeg over and over incurs process overhead, and if the video has distant keyframes (or in the worst case none at all) then the decoder must decode several frames for each invocation. I can do this one frame at a time: ffmpeg -i video.mp4 -ss 1.4 -frames 1 frame_1_4.jpgįfmpeg -i video.mp4 -ss 5.6 -frames 1 frame_5_6.jpgįfmpeg -i video.mp4 -ss 10.5 -frames 1 frame_10_5.jpg 1.4, 5.6, 10.5, etc.įor each timestamp, I want to extract the nearest frame and save it to a jpeg file.

Suppose I have a list of timestamps, e.g.

Extract frames ffmpeg python