The author proposes a lightweight masked auto-encoder model for efficient video anomaly detection, leveraging motion gradients and self-distillation to improve performance significantly.
Neural network models log-density function for video anomaly detection.
A self-supervised learning approach for video anomaly detection that leverages spatio-temporal coherence within video frames by predicting the order of shuffled patches in a video.
The proposed framework leverages image-text contrastive learning to extract local patterns that generalize to novel anomalies, and models the dynamics of these local patterns to effectively detect both spatial and temporal anomalies.
The proposed method transforms video anomaly detection into a deblurring process, focusing on motion features to achieve efficient anomaly detection and cross-domain generalization.
영상 이상 탐지를 위해 외관 이미지에 가우시안 블러를 적용하여 가짜 이상 데이터를 생성하고, 동작 특징을 기반으로 한 메모리 모듈을 활용하여 정상 동작 분포를 학습하고 이상 동작을 탐지한다.