Reproduce “ActionSwitch: Class-agnostic Detection of Simultaneous Actions in Streaming Videos” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hyolim Kang, Jeongseok Hyun, Joungbin An, Youngjae Yu, Seon Joo Kim
ActionSwitch, developed by researchers at Yonsei University, introduces the first class-agnostic framework for online temporal action localization, capable of detecting simultaneous actions in streaming videos. This framework utilizes a state-emitting online action detection model combined with a novel conservativeness loss to achieve state-of-the-art performance on datasets like FineAction, Epic-Kitchens 100, and THUMOS14.
Key insights
- Modeling action occurrences as state transitions via multiple independent switches allows for robust, class-agnostic detection of overlapping action instances, even for actions of the same class.
- The conservativeness loss effectively embeds a stability prior into the model's decision-making, significantly improving the precision and temporal continuity of detected action boundaries without complex architectural changes.
- Decoupling action localization from classification enhances flexibility, enabling the framework to handle scenarios with unknown classes or same-class overlaps, and supports future open-vocabulary applications.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.