Contrastive Self-Supervised Pre-Training for Video Quality Assessment.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Video Quality Assessment (VQA) is a small sample learning problem due to high annotation costs.
• Existing VQA models often use ImageNet pre-trained models, which are suboptimal for the different VQA domain.
• Limited scale of current VQA datasets hinders model development.

Purpose of the Study:

• To develop a self-supervised pre-training method for VQA.
• To leverage abundant unlabeled video data for feature representation learning.
• To improve the performance of VQA models by addressing data scarcity.

Main Methods:

• Proposed a self-supervised pre-training approach for VQA using contrastive learning.
• Introduced a distortion augmentation strategy to generate diverse video samples.
• Implemented contrastive learning to capture quality-aware information from video frames.
• Incorporated a distortion prediction task as an auxiliary objective.

Main Results:

• Achieved state-of-the-art results in self-supervised learning for VQA.
• Demonstrated significant performance benefits for existing learning-based VQA models.
• The pre-trained model effectively learns quality-aware feature representations.

Conclusions:

• Self-supervised pre-training with contrastive learning is highly effective for VQA.
• The proposed method overcomes limitations of small-scale datasets and suboptimal pre-trained models.
• The learned representations enhance downstream VQA tasks, advancing the field.