Untrimmed Action Anticipation

Egocentric action anticipation consists in predicting a future the camera wearer will perform from egocentric video. While task has recently attracted attention of research community, current approaches assume that input videos are “trimmed”, meaning short video sequence is sampled fixed time before beginning action. We argue that, despite recent advances field, trimmed limited applicability real-world scenarios where it important to deal with “untrimmed” inputs and cannot be assumed exact moment which begin known at test time. To overcome such limitations, we propose an untrimmed task, which, similarly temporal detection, assumes time, while still requiring predictions made actions actually take place. evaluation procedure for methods designed address this novel compare several baselines on EPIC-KITCHENS-100 dataset. Experiments show performance models very more required.