Industry surveillance is widely applicable in sectors like retail,
manufacturing, education, and smart cities, each presenting unique anomalies
requiring specialized detection. However, adapting anomaly detection models to
novel viewpoints within the same scenario poses challenges. Extending these
models to entirely new scenarios necessitates retraining or fine-tuning, a
process that can be time consuming. To address these challenges, we propose the
Scenario-Adaptive Anomaly Detection (SA2D) method, leveraging the few-shot
learning framework for faster adaptation of pre-trained models to new concepts.
Despite this approach, a significant challenge emerges from the absence of a
comprehensive dataset with diverse scenarios and camera views. In response, we
introduce the Multi-Scenario Anomaly Detection (MSAD) dataset, encompassing 14
distinct scenarios captured from various camera views. This real-world dataset
is the first high-resolution anomaly detection dataset, offering a solid
foundation for training superior models. MSAD includes diverse normal motion
patterns, incorporating challenging variations like different lighting and
weather conditions. Through experimentation, we validate the efficacy of SA2D,
particularly when trained on the MSAD dataset. Our results show that SA2D not
only excels under novel viewpoints within the same scenario but also
demonstrates competitive performance when faced with entirely new scenarios.
This highlights our method's potential in addressing challenges in detecting
anomalies across diverse and evolving surveillance scenarios.

我们提出了情景适应性异常检测（SA2D）方法，利用少样本学习框架对预训练模型进行快速适应新概念。为了解决数据集的缺乏问题，我们引入了多情景异常检测（MSAD）数据集，包括从各种摄像机角度捕获的 14 个不同情景，背景中包含挑战性的变化。通过实验证明了 SA2D 的有效性，尤其是在 MSAD 数据集上进行训练。这突出了我们的方法在检测不同和不断发展的监控情景中的异常的挑战中的潜力。