Introduction
Video Anomaly Detection (VAD) involves identifying rare and unpredictable events within surveillance footage. Traditionally, this "one-class learning" task relies on Generative Adversarial Networks (GANs) or Auto-Encoders (AEs) to model "normal" behavior. However, these architectures often produce blurry reconstructions and struggle with background noise, necessitating auxiliary semantic models like object detectors or optical flow estimators to achieve high performance. These "crutches" introduce significant computational overhead and potential points of failure.
The Feature Prediction Diffusion Model (FPDM) addresses these limitations by leveraging the superior generative and denoising capabilities of Diffusion Models (DMs). FPDM is the first framework to apply DMs to VAD by predicting frame features, eliminating the need for complex auxiliary tools.
Methodology
FPDM employs a two-stage diffusion pipeline operating on compressed feature maps extracted by an unsupervised 2D encoder. This approach avoids the high computational cost of pixel-level diffusion while maintaining rich semantic information.
Feature Prediction and Refinement
The core of FPDM consists of two specialized modules:
- Feature Prediction Diffusion Module (Motion): This module learns the distribution of normal motion by denoising a target feature map based on a sequence of previous "normal" features.
- Feature Refinement Diffusion Module (Appearance): To prevent the predicted features from becoming too generic, this module uses a cross-attention mechanism to inject appearance details from the original frame, focusing on the distribution of normal appearance.
A critical finding in the training process was that joint training of these modules led to poor performance. Instead, the authors utilized a staged training strategy, converging the prediction module before training the refinement module.
Experimental Results
FPDM was evaluated on major benchmarks including CUHK Avenue, ShanghaiTech, UCF-Crime, and UBnormal. The model achieved state-of-the-art results for image-based one-class VAD:
- ShanghaiTech: 78.6% AUC, a significant improvement over previous image-based methods.
- UCF-Crime: 74.7% AUC, outperforming several methods that utilize specialized 3D features.
To address the inherent latency of diffusion models, the authors implemented Denoising Diffusion Implicit Models (DDIM) sampling, resulting in a 20x speedup during inference compared to standard DDPM sampling.
Conclusion
FPDM demonstrates that Diffusion Models can effectively replace complex multi-model pipelines in video anomaly detection. By specializing in motion prediction and appearance refinement, the model achieves competitive performance using only basic 2D features. Future work may explore end-to-end training of the encoder and diffusion modules to further optimize feature alignment and detection accuracy.
Summary updated