Learning Fine-Grained Video Anomaly Detection from Normal Videos.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Video anomaly detection (VAD) methods are limited by the scarcity of high-quality annotated data, resulting in only video-level predictions.
• Existing methods for synthesizing pseudo video anomalies lack realism and fine-grained annotations, hindering real-world applicability.

Purpose of the Study:

• To develop a framework for unsupervised anomaly video generation using only normal videos.
• To enable fine-grained annotations (video, frame, and region levels) for improved VAD.
• To create a fine-grained VAD network capable of multi-level predictions.

Main Methods:

• Leveraging Vision-Language Models (VLMs) to generate structured textual descriptions of anomalies.
• Synthesizing abnormal video segments using VLMs based on generated textual descriptions.
• Developing a fine-grained VAD network trained on synthetically generated, annotated data.

Main Results:

• The proposed framework successfully generates unsupervised anomaly videos with high realism.
• The generated data provides fine-grained annotations, overcoming limitations of previous methods.
• The developed fine-grained VAD network achieves remarkable performance across video, frame, and region levels.

Conclusions:

• Unsupervised anomaly video generation using VLMs is a viable approach to address data scarcity in VAD.
• The proposed method enables the creation of detailed annotations, facilitating fine-grained VAD.
• This framework significantly advances the capabilities of VAD systems for real-world applications.