Reproduce “Benchmarking the Robustness of Temporal Action Detection Models Against Temporal Corruptions” 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

Runhao Zeng, Xiaoyong Chen, Jiaming Liang, Huisi Wu, Guangzhong Cao, Yong Guo

This research introduces the first comprehensive benchmark to evaluate the temporal robustness of Temporal Action Detection (TAD) models against realistic video corruptions. The study demonstrates that all leading TAD models are vulnerable to such corruptions, primarily affecting localization, and proposes a robust training strategy that improves both robustness and clean data performance.

Key insights

  • All seven evaluated Temporal Action Detection (TAD) models exhibit universal vulnerability to temporal corruptions, experiencing mAP drops ranging from 1.77% to 25.86% across benchmarks.
  • Performance degradation in TAD models due to temporal corruptions primarily results from increased localization errors rather than classification errors.
  • The proposed robust training strategy simultaneously improves model robustness against temporal corruptions and enhances performance on clean, uncorrupted data, challenging the traditional robustness-performance trade-off.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.