Anomalies are Streaming: Continual Learning for Weakly Supervised Video Anomaly Detection
Keywords: Weakly supervised video anomaly detection, continual learning
Abstract: Weakly supervised video anomaly detection (WSVAD) aims to locate frame-level anomalies with only video-level annotations provided. However, existing WSVAD methods struggle to adapt to real-world scenarios, where unseen anomalies are continuously introduced, thereby making the training of WSVAD essentially a process of continual learning. In this paper, we pioneer to explore the continual learning for weakly supervised video anomaly detection (CL-WSVAD), seeking to mitigate the catastrophic forgetting when the detection model learns new anomalies. We propose normality representation pre-training prior to continual learning, utilizing potential anomaly texts to guide the model in learning robust normality representations, which improves discrimination from potential incremental anomalies. Additionally, we introduce a mixed-up cross-modal alignment method to assist in adapting the pretrained model on CL-WSVAD. Subsequently, we propose a continual learning framework based on sequentially retaining the learnable text prompts for each type of anomaly, which effectively mitigates catastrophic forgetting. Experiments on our established CL-WSVAD benchmarks demonstrate the superiority of proposed method.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3776
Loading