Reproduce “Real-world Anomaly Detection in Surveillance 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

Waqas Sultani, Chen Chen, Mubarak Shah

Researchers from Information Technology University and the University of Central Florida developed a deep Multiple Instance Learning (MIL) framework for real-world anomaly detection in surveillance videos, leveraging weak video-level labels. The approach, enhanced with sparsity and temporal smoothness constraints, achieved an AUC of 75.41% on a newly introduced large-scale dataset, significantly outperforming prior methods and reducing false alarm rates to 1.9%.

Key insights

  • Weakly supervised Multiple Instance Learning successfully enables anomaly detection and localization within long, untrimmed videos without requiring precise temporal annotations.
  • Training models with both normal and anomalous video examples leads to more robust and discriminative anomaly detectors, evidenced by lower false alarm rates.
  • The incorporation of sparsity and temporal smoothness constraints in the loss function improves the model's ability to accurately pinpoint the short-duration anomalous segments.

Summary by alphaXiv.

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