To be useful in any project, we would like to approach real-time performance. Some algorithms implemented may not be able to achieve this (exhaustive search algorithms, for example), but we still set our goals high!
All tests computed for scikit-video v1.1.1 (historical data from 2018; updated benchmarks planned for a future release)
Minimum time from 10 trials, loading the 3 test videos in skvideo.datasets:
Method |
Time |
|---|---|
skvideo.io.vread (FFmpeg) |
0.718217 seconds |
skvideo.io.vread (LibAV) |
0.815005 seconds |
skvideo.io.vreader (FFmpeg) |
0.671952 seconds |
skvideo.io.vreader (LibAV) |
0.774765 seconds |
The fastest backend appears to be FFmpeg for both skvideo.io.vread and skvideo.io.vreader. Naturally, since skvideo.io.vreader uses a yield-based generator to supply frames, it is faster to use than skvideo.io.vread which allocates space then copies data frame-by-frame.
Using the default settings on the carphone_pristine.mp4 sequence, shape (120, 144, 176, 3)
Performance on block motion algorithms using skvideo.motion.blockMotion:
Method |
Time |
|---|---|
exhaustive |
43.946715 seconds |
3-step search |
17.558664 seconds |
“new” 3-step search |
32.340459 seconds |
simple and efficient 3-step search |
13.861776 seconds |
4-step search |
36.315580 seconds |
adaptive rood pattern search |
21.413136 seconds |
diamond search |
25.679036 seconds |
We can see that 3-step search is currently the fastest algorithm at 17.5 seconds. That comes out to 0.146 seconds per frame, for 144x176 sized frames. Clearly this is not realtime, which makes sense given that the core computations are in non-optimized python.