SSIM over MSE: A new perspective for video anomaly detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Pattern Recognition

Background:

• Current video anomaly detection models often fail to align with human perception due to reliance on Mean Squared Error (MSE).
• These models prioritize texture over shape, leading to poor interpretability and performance limitations.
• Existing methods struggle to accurately identify anomalies as perceived by the Human Visual System (HVS).

Purpose of the Study:

• To optimize video anomaly detection models by incorporating human visual relevance.
• To improve the alignment between model-detected anomalies and human perception.
• To enhance model interpretability and overall performance in detecting abnormal video patterns.

Main Methods:

• Introduction of a novel Structural Similarity Index (SSIM) based loss function.
• Development of a new anomaly score calculation method utilizing SSIM.
• Integration of a spatial-temporal enhancement block in 3D convolution (STE-3D) for improved feature capture.

Main Results:

• The SSIM-based loss emphasizes shape information over texture, enhancing model interpretability.
• The SSIM-based anomaly score aligns better with human visual perception.
• The STE-3D block effectively captures spatial-temporal features and compensates for SSIM loss limitations.
• Experimental validation on benchmarks like UCSD Ped1/Ped2, CUHK Avenue, and ShanghaiTech demonstrates significant performance improvements.

Conclusions:

• The proposed approach, integrating SSIM loss and STE-3D, effectively improves video anomaly detection.
• Aligning models with human visual perception leads to more interpretable and accurate anomaly detection.
• The methods are lightweight and seamlessly integrate with existing 3D convolutional models.