Skip to content

Files

Latest commit

ebeece3 · Dec 1, 2017

History

History

turn_codes

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
Sep 14, 2017
Sep 14, 2017
Dec 1, 2017
Sep 14, 2017
Sep 14, 2017
Sep 14, 2017
Sep 14, 2017
Sep 14, 2017
Sep 14, 2017
Sep 14, 2017
Sep 14, 2017
Sep 14, 2017
Sep 14, 2017
Sep 14, 2017
Sep 14, 2017

Sorry about the ugly coding style...

The model is built using TensorFlow 0.12.

The unit-level features should be saved with the name of (movie_name+".mp4"+"_"+str(swin_start)+"_"+str(swin_end)+".npy"), where movie_name is the name of the video, swin_start and swin_end are the start and end frame of the window. For example, video_test_0001495.mp4_10809.0_10825.0.npy. The unit-level feature should be extract every 8 frames (the unit size is 16 frames, so there will be 50% overlap.).

I uploaded my denseflow CNN unit features (size 16, overlap 50%) of THUMOS-14 to Google Drive: val set, test set.

Once you have the unit-level features, edit the feature path in main.py, and then just run python main.py. Best model is expected to be trained in about 10000 steps with current training samples.

The post_processing.py in test_results folder should be applied on the output test result file. After post processing, the pkl file can be evaluated by the eval program.