Reproduce “GTA-Crime: A Synthetic Dataset and Generation Framework for Fatal Violence Detection with Adversarial Snippet-Level Domain Adaptation” 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

Seongho Kim, Sejong Ryu, Hyoukjun You, Je Hyeong Hong

Researchers from Hanyang University developed GTA-CRIME, a synthetic dataset and generation framework for fatal violence detection, along with an adversarial snippet-level domain adaptation strategy. This approach improved fatal violence detection accuracy on real-world surveillance data, with one model (RTFM) increasing AUC from 85.43 to 87.27.

Key insights

  • Synthetic data with fine-grained frame-level annotations and diverse multi-view perspectives can serve as a valuable standalone resource for VAD, even for rare critical events.
  • Directly combining synthetic and real-world data without domain alignment can degrade performance for certain VAD models, particularly those not leveraging textual context.
  • Snippet-level adversarial domain adaptation effectively bridges the synthetic-to-real domain gap, enhancing real-world applicability without altering generalized pre-trained feature extractors.

Summary by alphaXiv.

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