CVD2014-A Database for Evaluating No-Reference Video Quality Assessment Algorithms.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Multimedia Signal Processing

Background:

• Existing video databases often use post-processing for distortions, not reflecting real-world camera acquisition.
• Subjective quality assessment is crucial for video quality evaluation but requires robust datasets.

Purpose of the Study:

• Introduce the CVD2014 database, a novel resource for video quality research.
• Provide observer-specific quality scores and descriptions to define quality dimensions.
• Evaluate the performance of video quality assessment algorithms.

Main Methods:

• Collected 234 videos from 78 different cameras, capturing complex, real-world distortions.
• Gathered observer-specific quality scores and open-ended descriptions for sharpness, graininess, color balance, darkness, and jerkiness.
• Developed a new performance measure accounting for observer variance to assess algorithms.

Main Results:

• The CVD2014 database offers a complex distortion space from real camera acquisition.
• Quality dimensions identified include sharpness, graininess, color balance, darkness, and jerkiness.
• Performance study indicates significant room for improvement in current video quality assessment algorithms.

Conclusions:

• CVD2014 provides a valuable, publicly available resource for advancing video quality research.
• Observer-specific data and real-world distortions are essential for accurate video quality assessment.
• Further development of video quality algorithms is needed to better predict subjective perception.