search

ikuinen / CMIN_moment_retrieval

Cross-Modal Interaction Networks for Query-Based Moment Retrieval in Videos
86 stars 20 forks source link

How to get the video C3D features' vectors datasets #2

Closed wangsu1999 closed 5 years ago

wangsu1999 commented 5 years ago

Hello,

Thanks for your sharing. But I have something trouble to get the datasets of the video C3D features' vectors ,which is fit for this codes.Could you tell me about the steps to get the C3D features or share your C3D features?

Best

ikuinen commented 5 years ago

Hello,

Thanks for your sharing. But I have something trouble to get the datasets of the video C3D features' vectors ,which is fit for this codes.Could you tell me about the steps to get the C3D features or share your C3D features?

Best

The C3D features of ActivityCaptions can be downloaded here, and the way of feature extraction is presented here:

The publicly available pre-trained C3D model which has a temporal resolution of 16 frames was used to extract frame based features. This network was not fine-tuned on our data. We reduce the dimensionality of the activations from the second fully-connected layer (fc7) of our visual encoder from 4096 to 500 dimensions using PCA. The C3D features were extracted every 8 frames.

This is exactly the way we apply to extract visual features of videos in TACoS. Hope this may help you!

wangsu1999 commented 5 years ago

Thank you. It is very helpful for me.