search

yytzsy / SCDM

Code for the paper: Semantic Conditioned Dynamic Modulation for Temporal Sentence Grounding in Videos
68 stars 14 forks source link

I3D extractor #3

Closed Tanwey closed 4 years ago

Tanwey commented 4 years ago

Hi Yitian,

I need to more precise I3D features(higher fps).Could you please share the codes for I3D extractor? Thanks

TianyuLee commented 4 years ago

I have a similar question: Are the I3D features and C3D features fine-tuned in (charades,activitynet) dataset? or just trained in (sports-1m , kinetics) dataset? thanks!

yytzsy commented 4 years ago

We do not finetune the feature extraction network in any other datasets, and just use the pretrained model on sports-1m to get the video features. For C3D feature extraction, we follow https://github.com/yyuanad/Pytorch_C3D_Feature_Extractor, and for I3D feature extraction, please refer https://github.com/JaywongWang/I3D-Feature-Extractor

TianyuLee commented 4 years ago

OK, thank you and your code very much!

Tanwey commented 4 years ago

Thanks for your sharing!

Sy-Zhang commented 4 years ago

For I3D extraction of Charades, I got same result with the "rgb600" checkpoint (pretrained on Kinetics 600) instead of the default "rgb_imagenet" checkpoint (pretrained on Kinetics 400). Also, All videos are decoded at 16 fps.