About the extracted features

[cpsxhao]

Hi,

Thank you very much for your great code. May I ask if you would provide the extracted and pre-processed features (the .npz files) for datasets such as COCO, VCR, MSVTT?

Thank you very much!

[YehLi]

I have uploaded the msvd and msr-vtt features to the msvd_dataset/msrvtt_dataset folder (https://drive.google.com/drive/folders/1vx9n7tAIt8su0y_3tsPJGvMPBMm8JLCZ?usp=sharing).

For MSCOCO dataset, please down the bottom up features from https://github.com/peteanderson80/bottom-up-attention and convert them into npz format with tools/create_feats.py.

[cpsxhao]

Hi, thanks very much for your help! BTW, may I ask where can I find the "mirror" features, or how can I convert the features in the bottom-up-attention repo into "mirror" npz-format features?

[YehLi]

The mirror features are extracted with the mirror images by https://github.com/peteanderson80/bottom-up-attention/blob/master/tools/generate_tsv.py.

[HanielF]

@YehLi Hi~
I notice that a video contains several frames and each frame can be used to extract up to 100 bboxes and roi features. Does the video npz-format features are assembled by concatting frames features？
The generate_tsv.py script only works for single image other than multiple frames.