FADMB: Fully attention-based dual memory bank network for weakly supervised video anomaly detection

Zhiming Luo, Shuheng Huang, Kun Yang, Jianzhe Gao, Shaozi Li

Abstract: Highlights•FADMB Introduction: The proposed FADMB replaces the Top-k selection mechanism with an attention-based paradigm for updating memory banks, overcoming limitations and enhancing performance in weakly supervised video anomaly detection (VAD).•Hybrid Encoder: FADMB introduces a Hybrid Encoder that integrates both local and global temporal information using an Inception-like LSTM for local context and multi-head self-attention for global context, improving feature representation.•State-of-the-art Performance: Extensive experiments demonstrate that FADMB outperforms other methods, achieving top results on benchmark datasets like XD-Violence and UCF-Crime.•Robust Memory Bank Update: The attention-based memory bank learning mechanism avoids information loss, bias, and hyperparameter tuning challenges associated with Top-k selection, leading to more comprehensive anomaly detection.•Frame-level Consistency: Qualitative visualizations show that FADMB’s results align closely with frame-level anomaly annotations, providing consistent and accurate anomaly detection.

External IDs:dblp:journals/pr/LuoHYGL26