Perceptual Geometry Distortion Assessment of Compressed 3D Meshes.

Area of Science:

• Computer Vision
• Computer Graphics
• Signal Processing

Background:

• Evaluating perceptual quality in 3D mesh compression, especially for Video-based Dynamic Mesh Coding (V-DMC), faces challenges due to limited subject-rated datasets and high computational demands for full decoding and feature extraction.
• Existing methods often require complete mesh decoding or complex feature extraction, hindering real-time applications.

Purpose of the Study:

• To address the limitations in V-DMC perceptual quality assessment.
• To introduce a novel dataset and a computationally efficient no-reference model for V-DMC quality evaluation.

Main Methods:

• A new V-DMC distortion dataset was created, featuring 16 original meshes and 400 compressed variants.
• Subjective quality assessment was performed using the Double Stimulus Impairment Scale (DSIS) with 30 participants to gather Mean Opinion Scores (MOS).
• A novel no-reference, bitstream-layer model, streamMQ, was developed to predict perceptual quality by extracting features directly from the compressed bitstream without full decoding.

Main Results:

• The streamMQ model demonstrates highly competitive perceptual quality assessment performance compared to state-of-the-art methods.
• streamMQ achieves this performance with significantly reduced computational and storage costs.
• The proposed method facilitates real-time and low-storage application environments for V-DMC.

Conclusions:

• The developed V-DMC dataset and streamMQ model offer a valuable resource for advancing perceptual quality assessment in 3D mesh compression.
• streamMQ provides an efficient and effective solution for real-time V-DMC quality evaluation, overcoming previous computational barriers.
• The public availability of the dataset and code (https://github.com/HFL01/QDU-GDM) will foster further research and development in this area.