gistfile1.txt

https://github.com/kkroening/ffmpeg-python/blob/master/examples/ffmpeg-numpy.ipynb
pip install ffmpeg-python

https://github.com/quanhua92/human-pose-estimation-opencv.git
ffmpeg -i "in.mp4"  -s 240x135 -vf fps=1 %d.jpg

dir -recurse -path data -include *.jpg -Name | %{echo "python openpose.py --input $_ --output output\sized\$_ --thr .1" }
find ".\data\" --name "data\*.jpg" --exec "python openpose.py --input {} --output output\{} --thr .1" \;


git clone https://github.com/pjreddie/darknet
cd darknet
make

# include/darknet.h:780:11: error: unknown type name 'clock_t'; did you mean 'clockid_t'?
https://github.com/pjreddie/darknet/issues/217
make -nogpu # works; resolves clock_t issue above.
cd cfg; wget https://pjreddie.com/media/files/yolov2.weights


module 'tensorflow' has no attribute 'InteractiveSession'
change sess = tf.InteractiveSession() to sess=tf.compat.v1.InteractiveSession()


----
Came across this question, so here's a quick comparison. Compare these two different ways to extract one frame per minute from a video 38m07s long:

time ffmpeg -i input.mp4 -filter:v fps=fps=1/60 ffmpeg_%0d.bmp
1m36.029s

This takes long because ffmpeg parses the entire video file to get the desired frames.

time for i in {0..39} ; do ffmpeg -accurate_seek -ss `echo $i*60.0 | bc` -i input.mp4   -frames:v 1 period_down_$i.bmp ; done
0m4.689s

This is about 20 times faster. We use fast seeking to go to the desired time index and extract a frame, then call ffmpeg several times for every time index. Note that -accurate_seek is the default , and make sure you add -ss before the input video -i option.

Note that it's better to use -filter:v -fps=fps=... instead of -r as the latter may be inaccurate. Although the ticket is marked as fixed, I still did experience some issues, so better play it safe.