Global-Context Aggregated Intra Prediction Network for Depth Video Coding

Abstract

The huge data volume of 3D video brings great challenges to the storage and deployment of existing multimedia systems. To reduce the redundancy of 3D video, this paper proposes a global-context aggregated intra prediction network (GCAIP-Net) to improve the performance of depth video coding. In contrast with previous studies only exploiting the local context in reference pixels, the proposed GCAIP-Net takes the advantage of the transformer to explore the global context for more accurate prediction. Specifically, a transformer-based prediction module is designed to generate predicted features by modeling long-range dependencies in the spatial domain. The proposed method is integrated into the 3D-HEVC reference software HTM16.2, and extensive experiments show that the proposed method achieves 3.62% BD-rate reduction compared with 3D-HEVC under all-intra configuration.