Fast 3D-HEVC Depth Map Coding Method Based on Spatio-Temporal Correlation and a Two-Stage Mode Decision Framework

Area of Science:

• Computer Vision
• Video Compression
• Machine Learning

Background:

• Efficient intra-mode decision is crucial for 3D-HEVC performance.
• Current methods using texture or ML have limitations in leveraging spatio-temporal correlations and handling complex regions.
• Deterministic classifiers lack reliability in edge mutation or intricate areas.

Purpose of the Study:

• To develop a fast intra-mode decision algorithm for depth maps.
• To improve the efficiency and accuracy of mode selection in 3D-HEVC encoding.
• To balance encoding complexity and rate-distortion performance.

Main Methods:

• A two-stage algorithm integrating naive Bayes probability estimation and fuzzy support vector machine (FSVM).
• Stage 1: Spatio-temporal prior modeling to confine candidate mode space.
• Stage 2: FSVM for enhanced decision accuracy in low-confidence regions.

Main Results:

• Reduced average encoding time by 52.30%.
• Achieved only a 0.68% increase in BDBR (Best-fit Distortion-Bitrate).
• Demonstrated stable performance across diverse test sequences and resolutions.

Conclusions:

• The proposed algorithm effectively reduces computational complexity while maintaining rate-distortion performance.
• It offers a significant improvement in encoding efficiency for depth map processing in 3D-HEVC.
• The method provides a robust and universally applicable solution for depth map intra-mode decision.