V3C1-Pseudo-Caption (V3C1-PC) is an auto-generated video description dataset for (pre-)training video-text matching models. Pseudo captions for a given video in the V3C1 collection were generated as follows. We used BLIP to generate a caption for each sampled frame. An n-frame video will have n captions. We removed duplicate caption and then used CLIP to rank the remaining captions in terms of their cross-modality similarity to the video. The top-3 ranked captions were preserved as the video’s pseudo captions. V3C1-PC consists of 436,204 captions for 219,531 video shots.
- Caption data (20MB): Google Drive Link
The dataset was developed during our participation (TeamID: RUCMM) in the TRECVID 2022 Ad-hoc Video Search (AVS) task.
@inproceedings{tv22-rucmm,
title = {Renmin {U}niversity of {C}hina at {TRECVID} 2022: Improving Video Search by Feature Fusion and Negation Understanding},
author = {Xirong Li and Aozhu Chen and Ziyue Wang and Fan Hu and Kaibin Tian and Xinru Chen and Chengbo Dong},
booktitle = {TRECVID 2022 Workshop},
year = {2022},
}