대조 학습을 통한 비디오 이상 탐지를 위한 시각적 특징과 텍스트 특징 정렬

다국어 초록 (Multilingual Abstract)

Video anomaly detection has witnessed significant advancement since the development of the multiple instance learning (MIL) approach [1], and recently is expanding to incorporate both visual and textual features of videos [3]. The main idea behind this is that text features also contain frame-specific information that can complement visual features. In this work, we aim to enhance video anomaly detection by introducing a contrastive approach to robustly align visual features and textual features. For this purpose, we propose a loss function to increase the similarity between frame-level visual features and textual features [5]. Experimental results demonstrate that this approach is effective when applied to existing algorithms.