Completely Blind Quality Assessment of User Generated Video Content.

Area of Science:

• Computer Vision
• Computational Neuroscience
• Signal Processing

Background:

• Blind video quality assessment (BVQA) for user-generated content (UGC) is challenging due to lack of ground truth and undefined distortion models.
• Existing methods often rely on training data or specific distortion assumptions, limiting their applicability to diverse UGC.

Purpose of the Study:

• To propose a novel, training-free BVQA method for UGC.
• To leverage computational neuroscience principles of the human visual system (HVS) for quality prediction.
• To develop a robust and explainable video quality assessment metric.

Main Methods:

• Inspired by the hypothesis that the HVS transforms natural videos to follow a straighter temporal trajectory in the perceptual domain.
• Modeled the lateral geniculate nucleus (LGN) and V1 regions using bandpass filters to validate the perceptual straightening hypothesis.
• Quantified temporal quality by measuring the loss of straightness (curvature) in transformed video representations.
• Combined the temporal quality measure with a spatial quality metric to create the STraightness Evaluation Metric (STEM).

Main Results:

• Extensive empirical evidence validated the hypothesis that video distortions increase curvature in HVS-transformed representations.
• The proposed temporal quality measure demonstrated acceptable performance independently.
• The developed STraightness Evaluation Metric (STEM) achieved state-of-the-art performance across five UGC VQA datasets.
• STEM proved to be completely blind (training-free), generalizable, explainable, and simple to implement.

Conclusions:

• The perceptual straightening hypothesis provides a viable foundation for BVQA of UGC.
• STEM offers a significant advancement in blind video quality assessment, outperforming existing methods.
• The training-free, explainable, and generalizable nature of STEM makes it highly practical for real-world applications.