Reproduce “Anomaly Detection for People with Visual Impairments Using an Egocentric 360-Degree Camera” 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
Inpyo Song, Sanghyeon Lee, Minjun Joo, Jangwon Lee
A novel system using an egocentric 360-degree camera detects subtle, short-duration physical safety anomalies for people with visual impairments, achieving 86.00% AUC-ROC on a new VIEW360 dataset and accurately predicting the anomaly's direction with 75.04% accuracy. The Frame and Direction Prediction Network (FDPN) employs saliency-driven masking and coarse-to-fine learning to enable precise, frame-level anomaly detection and directional localization.
Key insights
- Saliency-driven masking effectively guides the network to visually significant areas in wide 360-degree views, improving anomaly detection performance over general object detection.
- The coarse-to-fine learning approach successfully derives precise frame-level anomaly pseudo-labels from video-level weak supervision, enabling accurate detection of subtle, short-duration events.
- Combining image features with saliency information is crucial for accurately predicting the direction of anomalous events, providing actionable information for users.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.