Where can we find the find the .fc6-1 files? about anomalydetectioncvpr2018

[WaqasSultani]

Hi Mourad,
Sorry, I don't have those features. However, you don't have to use C3D. It's just the video features that we used in our paper. You can use I3D (https://github.com/deepmind/kinetics-i3d/blob/master/i3d.py) or Multifiber network (https://github.com/cypw/PyTorch-MFNet). I believe you should be able to compute them.
I got an email from someone that he used I3D instead of C3D and got better results.
I believe I3D or Multifiber network could be easily run on freely available GPU (Google Collab).
Good luck with your MS thesis.

Best,

[mourad1081]

Hello, thank you very much for your answer. I'll dig further on what you suggest me.

Thanks,

[Anirudh58]

Hi @mourad1081

Just a suggestion if you are still trying to use C3D, I had a similar running out of memory problem since I did not have any powerful GPU.
You can however, evade this issue by explicitly mentioning that you want to use only CPU.
The script inside c3d_sport1m_feature_extraction_video.sh should be changed as follows.

GLOG_logtosterr=1 ../../build/tools/extract_image_features.bin prototxt/c3d_sport1m_feature_extractor_video.prototxt conv3d_deepnetA_sport1m_iter_1900000 -1 10 100000 prototxt/output_video_list.txt fc6-1

The '-1' mentions that you want to use only CPU.

It will however be very slow, obviously. Took me almost 4 days to extract FC6 features from 1600 videos.

[mourad1081]

Hello @Anirudh58 ,

Thank you for your suggestion. I have tried that workaround some weeks ago but it was too slow (I had to let my computer run for a hundred days 24/7).

Fortunately, a friend bought me a GTX 1080 and everything works fine regarding C3D. It took roughly one hour to process several hours of video.