About hdf5 files about frameglimpses

[buvaev]

@Patrickcxt , this might be helpful
https://github.com/duguyue100/neuralfeature
https://github.com/garythung/torch-lrcn

[cloudandcat]

@bityangke I have prepared for data files，but when i trained it ， i got some problems.

[wandering007]

@cloudandcat I found that most of UCF101 videos are less than 10 seconds. If they are sampled at 5fps, we cannot get 50 frames...

[wandering007]

@bityangke Padding what frames? last one? I thought about it, but finally gave up. I think that it would bring incorrect temporal info.

[wangxiao5791509]

@bityangke Hi, would you please tell me how to pre-process the UCF101 dataset ? i.e. how to transform the *.avi video to hdf5 files ??? Thank you.

[Gandhi-Sagar]

Does anyone know how to generate meta file with THUMOS-14? The training data contains the trimmed videos. So the ground truth will always be 0 and last_frame_no, right? Also, do we need to use background data? Could we collectively work towards closing the issue? May be convey what we know? I think I have figured out how to generate hdf5 files. Will update based on a success..

[June01]

@Gandhi-Sagar Could you share your idea about how to generate hdf5 files? If needed, I can also help to work with meta data or others that haven't finished.
@syyeung Could you please give some reference ahout how to generate the data as you write? It will be a great help in this problem if you provide them all. Thank you very much~