DETAILS about using scripts/extract_optical_flow.sh about temporal-segment-networks HOT 12 CLOSED

@hawklucky hello, I meet the same problem as you. Could you tell me which py file you changed to fix the problem on removing the space between the sys cmd? Thank you.

I have found the problem:
in the tsn/tools/build_of.py,
when using the CommandLineParser, the original line is -f {} -x {} -y {} -i {} -b 20 -t 1 -d {} -s 1 -o {} -w {} -h {}, however, in the new verision opencv, it should be written in -f={} -x={} -y={} -i={} -b=20 -t=1 -d={} -s=1 -o={} -w={} -h={}.

If anyone meet the problem say "Cannot open video stream # "true" for optical flow extraction", this solution may solve your question.

from temporal-segment-networks.

Hi @hawklucky, can you share your modifications in detail? I have similar problem.

Updated: My problem is solved :)

from temporal-segment-networks.

This log message means that it cannot open the video file. It may be due
to some video files missing. Or it may be caused by OpenCV not having
VideoIO support.

On Sunday, August 28, 2016, Hawk Wang [email protected] wrote:

Thanks for your excellent work and code!
as above, when i using the dense flow to extract images and opt flow, the
FATAL occure

2016-08-28 21:16:28,702 FATAL [default] Check failed:
[video_stream.isOpened()] Cannot open video stream "true" for optical flow
extraction.
2016-08-28 21:16:28,702 WARN [default] Aborting application. Reason: Fatal
log at [/home/wanghao/my_project/temporal-segment-networks/lib/
dense_flow/src/dense_flow_gpu.cpp:17]

my python is 2.7.6, i think maybe the reason of function pool.map, can you
share some thoughts about it?

Thanks a lot!

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
#6, or mute
the thread
https://github.com/notifications/unsubscribe-auth/AGg4d_sJkIoAOqMoD4jMmQRM3hlNd_dhks5qkYrUgaJpZM4Ju7E2
.

Best regards,

Yuanjun

from temporal-segment-networks.

thanks a lot, just fixed by remove the space between the sys cmd in python

from temporal-segment-networks.

I train the ucf101 successfully with the data size 240_320_3, but failed when the hmdb51 the data size is different, I already realize that. By the way, May I ask that when I extract the image and flow, should I set the size 240_320 or 256_340 ? Thanks!

from temporal-segment-networks.

Setting 256*340 will be better. But both sizes should work.

from temporal-segment-networks.

@hawklucky
BTW, you mentioned "failed". What exactly does "failed" mean?
Also, did you update to the latest TSN codebase?

from temporal-segment-networks.

@yjxiong I use the branch of opencv-3.1 in dense_flow, it's no resize operation before calc flow and also no size args to use, so when i extract the image and flow, it's just the size of the origin data size, so it's failed because the size in CNN.

from temporal-segment-networks.

Alright, if it is the case, I may suggest a workaround for this.

In the prototxt file, you can ask the data layer to pre-scale the input image before augmentation. To do this, set

new_width: 340
new_height: 256

in video_data_param of the VideoData layers.

from temporal-segment-networks.

Yes, it's convenient to solve the problem by using new_width and new_height, thanks.

from temporal-segment-networks.

@plaovem
you are right, thank you very much!

from temporal-segment-networks.

@hawklucky I'm working with TSN. I having the same problem as you. the empty file generated? How did you solve it?

from temporal-segment-networks.