Chemical Sequence Verification Dataset

ShanghaiTech University

[Download(OneDrive)] [Download(BaiduNetDisk, extraction code: 9uyk)]

Introduction

This dataset is proposed to support the Sequence Verification task which aims to distinguish positive video pairs performing the same action sequence from negative ones with step-level transformations but still conducting the same task. The above figure demonstrates specific examples of our task: there are 3 step-sequences (seq1, seq2, seq3) included while seq1 and seq2 contain identical steps in the same order which leads (seq1, seq2) to be a matching pair; while seq2 and seq3 are not only different in the order of steps, but also in the number of steps, making (seq2, seq3) a dismatching pair.

Dataset Format

The CSV dataset contains 960,458 frames of over 1940 videos across 70 different kinds of procedures. On average, each video lasts 20.58 seconds, contains 495.85 frames, and consists of 9.53 steps.


As shown in the above figure, CSV includes 18 atomic-level actions with different frequencies, among which take and put are the two most common actions. This makes sense since taking up or putting down something are extremely common in reality. The videos’ length varies from 5.63s to 58.43s due to the diversity in complexity among procedures and individual differences of participants, such as movement habits, the memory of the step sequence as well as familiarity with the operations. By interacting one action with different objects, we have 106 steps in total (listed in the figure below).


Label info:
  • The labels follow the from of A.B where A indicates the task and B indicates one of the procedures of task A.
  • The realistic annotation for labels can be found in label_bank.json in the github repo .
  • Citation

    If you find this useful, please cite our work as follows:
    @article{qian2021svip,
      title={SVIP: Sequence VerIfication for Procedures in Videos},
      author={Qian, Yicheng and Luo, Weixin and Lian, Dongze and Tang, Xu and Zhao, Peilin and Gao, Shenghua},
      journal={arXiv preprint arXiv:2112.06447},
      year={2021}
    }