Efficient Inter-Prediction Method Using Reference Frame Accumulation for MPEG G-PCC

This paper presents a novel hybrid reference frame generation method to enhance the inter-prediction performance of geometry-based point cloud compression (G-PCC), a recent point cloud coding standard developed by the moving picture experts group (MPEG) that currently employs a single past reference frame for prediction. However, this single-reference approach may not fully capture the temporal and spatial correlations between frames, potentially limiting prediction performance. To address this limitation, we propose a hybrid mode selection scheme that chooses the suitable reference frames from a set of candidate frames, thereby leveraging the spatial and temporal correlations among consecutive past frames more effectively and improving the coding efficiency. We introduce and evaluate three coding modes for the hybrid reference frame concept within the G-PCC test model: 1) entropy estimation-based mode selection, which selects the reference frame(s) that minimize the estimated entropy during inter-frame coding; 2) translation-based mode selection, which selects the reference frame(s) requiring the least translation to align with the current frame; and 3) lightweight entropy estimation-based mode selection, which minimizes entropy under a predefined translation constraint. The experimental results demonstrate that all proposed techniques outperform the current G-PCC standard; entropy-estimation-based mode selection achieves an average improvement of 0.86% in geometric coding efficiency and 0.81% overall. Translation-based mode selection improved these metrics by 0.59% and 0.56%, respectively, and the lightweight entropy estimation approach consistently yielded a gain of 0.7%. In specific driving scenarios, all proposed techniques showed significant improvements, reaching approximately 2.62% to 2.86% in geometric coding efficiency.