Tragedy Plus Time
Capturing Unintended Human Activities from Weakly-labeled Videos

Arnav Chakravarthy

achakr37@asu.edu

Zhiyuan Fang

zy.fang@asu.edu

Yezhou Yang

yz.yang@asu.edu

We curate W-Oops, an augmented video dataset, built upon Oops, containing high quality video level labels for the goal-directed as well as unintentional action.

We propose an attention mechanism based method that highlights relevant temporal regions of the video important to a classification task when inferring the goal-directed and unintentional action while also ensuring their temporal ordering without using ground truth temporal labels during model training.

Finally, we provide in-depth and comprehensive experimental analysis, which validates the effectiveness of our method compared to competitive weakly supervised action localization methods on W-Oops.

We introduce W-Oops, a dataset built upon Oops*, a new model and comprehensive analysis

See it in Action!

Reference

Chakravarthy, A., Fang, Z., & Yang, Y. (2022). Tragedy Plus Time: Capturing Unintended Human Activities from Weakly-labeled Videos.
Bibtex