Reproduce “TF-CADE: Foreground-Concentrated Text-Video Alignment for Zero-Shot Temporal Action Detection” 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
Yearang Lee, Ho-Joong Kim, Seong-Whan Lee
Researchers from Korea University developed TF-CADE, a framework for Zero-Shot Temporal Action Detection that explicitly aligns textual information with action-relevant foreground regions in videos. TF-CADE improved average mAP on THUMOS14 (50%-50% split) from 16.0% to 21.1% and demonstrated stronger cross-dataset generalization by reducing background interference during text-video alignment.
Key insights
- Background information significantly interferes with effective text-video alignment in ZSTAD, leading to text-irrelevant predictions and poor class discriminability.
- Explicitly concentrating textual alignment on action-relevant foreground regions is crucial for improving semantic consistency and inter-class discriminability in zero-shot action detection.
- Combining foreground-weighted video embeddings with certainty-based confidence re-weighting enhances both classification and localization accuracy by providing a video-level prior.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.