High-Quality Dynamic Human Novel View Synthesis Based on Multiview Video

High-quality novel view synthesis (NVS) of dynamic humans is crucial for immersive applications such as VR/AR gaming and sports broadcasting. Although 3D Gaussian Splatting (3DGS) has emerged as a promising technique for NVS with superior training efficiency and rendering quality, its practical use in high-resolution dynamic human scenarios is limited by significant storage consumption and training time overhead. To address these challenges, we propose a coarse-to-fine training pipeline that incorporates attribute pruning, motion prediction, and neural encoding to substantially accelerate the training process. Furthermore, we introduce a dynamic 3DGS model compression algorithm that reduces spatio-temporal redundancy across the Gaussian sequence, leading to a significant decrease in storage requirements. Our method reduces the average training time per frame from 3 minutes to just 30 seconds, while achieving a 26× compression ratio in model storage compared to vanilla 3DGS. These improvements make 3DGS a viable solution for high-quality dynamic human NVS in real-world applications.