search

Finspire13 / CMCS-Temporal-Action-Localization

Completeness Modeling and Context Separation for Weakly Supervised Temporal Action Localization (CVPR2019)
MIT License
150 stars 17 forks source link

about the size of the pre-extracted feature #4

Closed yangwf1 closed 4 years ago

yangwf1 commented 4 years ago

hi, thanks for your sharing of the extracted features. But i'm curious why the size of the extracted features is so large. since the validation set of THUMOS14 is about 80G, and the feature here for it is also about 80G

Finspire13 commented 4 years ago

It is because ten-crop data augmentation is used (10x file size). And there are two type of features, i.e., I3D and UntrimmedNet.

Duplicate of #1

yangwf1 commented 4 years ago

@Finspire13 Thanks, really appreciate it. And i have one more question, is the data augmentation important or not? Is it convenient for you to report your result on the resized feature.

Finspire13 commented 4 years ago

@yangwf1 Ten-crop augmentation is about 2 mAP higher than the resized feature (cant find the precise number now). Note that the results of ablation studies in the paper are all on the ten-cropped features.