Title: Spatio-Temporal Correlation Guided Geometric Partitioning for Versatile Video Coding

Abstract:

Within the hybrid video coding framework, geometric partitioning has garnered significant interest due to its exceptional ability to describe motion fields. Nevertheless, the current geometric partitioning (GEO) standard in Versatile Video Coding (VVC) imposes a substantial overhead for signaling side information, thereby restricting coding efficiency. To address this challenge, we introduce a spatio-temporal correlation guided geometric partitioning (STGEO) scheme designed to efficiently characterize object information within the video coding motion field. This approach aims to reduce the bit consumption associated with side information signaling, encompassing both partitioning modes and motion data.

We begin by statistically analyzing the traits of partitioning mode decisions and motion vector selections. Leveraging the observed spatio-temporal correlations, we develop a mode prediction and coding technique intended to minimize the overhead required to represent this side information. The core strategy involves predicting STGEO modes and motion candidates with high selection probabilities, which then guide the entropy coding process by assigning fewer bits to these high-likelihood modes and motion candidates. Specifically, high-probability STGEO modes are predicted using edge data and the historical modes of neighboring STGEO-coded blocks. The associated motion information is encoded via an index within a merge candidate list, which is adaptively deduced based on merge candidate selection probabilities derived from offline training.

Experimental results demonstrate that the proposed method yields average bit-rate savings of 0.95% and 1.98% compared to VTM-8.0 without GEO, under Random Access and Low-Delay B configurations, respectively.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC