BiGS: Bi-Anchor Gaussian Splatting for Dynamic Video Photorealistic Reconstruction

Reconstructing photorealistic dynamic scenes from videos remains a challenging problem due to complex object motions, occlusions, and limited observations. Recent advances in dynamic Gaussian splatting either rely on deformation fields, which are compact but often fail under irregular motion, or per-frame trajectory optimization, which captures motion explicitly but incurs high redundancy. In this work, we introduce Bi-Anchor Gaussian Splatting (BiGS), a novel anchor-conditioned framework that combines the advantages of both paradigms. We separate the scene into static and dynamic Gaussians, where static Gaussians are shared across the sequence and dynamic Gaussians are explicitly optimized at sparse anchor frames. For non-anchor frames, we propose a bi-anchor motion field that predicts Gaussian transformations conditioned on two neighboring anchors, ensuring locally coherent and flexible dynamic modeling. To resolve inconsistencies between anchor proposals, we further design an opacity-gated fusion module that adaptively regulates per-Gaussian contributions. In addition, a priority-driven densification strategy introduces new Gaussians in regions with high photometric error, strong gradients, or insufficient observations, significantly improving detail reconstruction. We evaluate BiGS on DyCheck, Nvidia-long, and Neu3D datasets, and achieve competitive or superior reconstruction quality across PSNR, SSIM, and LPIPS compared with state-of-the-art NeRF- and Gaussian-based dynamic reconstruction methods.