Jingwen Wang

Sorry for the late response. What is the "init_from" you pass to ? It needs to point to the checkpoint abolute path.

I am not sure what caused this error. Have you checked if the following file exists: densevid_eval-master/coco-caption/pycocoevalcap/meteor/data/paraphrase-en.gz ? If NOT, please manually download it from https://github.com/tylin/coco-caption/tree/3f0fe9b819c0ea881a56441e4de1146924a394eb/pycocoevalcap/meteor/data .

As it is.

I am not sure what caused this error.

> Hello JaywongWang, > > As their is not direct path given in your repository "densevid_eval-master/coco-caption/pycocoevalcap/meteor/data/" > So do we need to create data folder manually and then need to...

I3D requires stacking multiple frames as input. If you want dense features, you need to densely sample 64 frames as input with one frame as stride.

However, we still can not guarrantee n_frames*1024 featrures.

Sorry for the late reply. I used tensorflow==1.14.0

I am really sorry for missing your question. I think the videos for action detection should be included in the 13320/1010/1574 video split. First, you need to download http://crcv.ucf.edu/THUMOS14/Validation_set/TH14_Temporal_annotations_validation.zip and...

@lijiannuist How can we train the progressive anchor loss? Should we directly set it to be True and train the whole network? (face + head + body)