Few-shot fine-tuning with auxiliary tasks for video anomaly detection

Anomaly detection in surveillance videos aims to identify video frames that exhibit unexpected behavior. Most existing methods follow an unsupervised setup, training with normal videos and testing with videos from the same scene. However, in real-world deployments, the performance of existing models significantly degrades when faced with unseen scenes. To address this issue, we introduce the auxiliary tasks of segmentation and optical flow estimation into the fine-tuning process, proposing a novel Segmentation and Optical Flow Fine-tuning (SOFF) framework. This framework enables the existing models to adapt to new scenes with only a few samples for fine-tuning. To integrate these auxiliary tasks, we design a Segmentation and Flow Output Network (SFO-Net). SFO-Net enhances fine-tuning performance in unseen scenes by extracting rich shape and motion information through the execution of auxiliary tasks during the fine-tuning process. Additionally, SFO-Net can be flexibly cascaded with existing models that output images to form the SOFF framework. Experiments on multiple datasets demonstrate that our framework improves the performance of existing models when faced unseen scenes through few-shot scenes fine-tuning and achieves competitive performance.