ADC-GS: Anchor-Driven Deformable and Compressed Gaussian Splatting for Dynamic Scene Reconstruction

TL;DR

ADC-GS addresses inefficiencies in 4D Gaussian Splatting by organizing Gaussians into an anchor-based canonical space and deforming them via a coarse-to-fine, anchor-driven pipeline. It couples this with a multi-dimension entropy model and adaptive rate-distortion optimization, together with a temporal-significance–guided anchor refinement to balance fidelity and bitrate. The method delivers substantial storage reductions (up to ~$204\times$ on Neu3D) and fast rendering (up to 3× faster than per-Gaussian deformation approaches) while maintaining high rendering quality on HyperNeRF and Neu3D. These results demonstrate practical advantages for real-time transmission and scalable dynamic scene capture, with code released for reproducibility.

Abstract

Existing 4D Gaussian Splatting methods rely on per-Gaussian deformation from a canonical space to target frames, which overlooks redundancy among adjacent Gaussian primitives and results in suboptimal performance. To address this limitation, we propose Anchor-Driven Deformable and Compressed Gaussian Splatting (ADC-GS), a compact and efficient representation for dynamic scene reconstruction. Specifically, ADC-GS organizes Gaussian primitives into an anchor-based structure within the canonical space, enhanced by a temporal significance-based anchor refinement strategy. To reduce deformation redundancy, ADC-GS introduces a hierarchical coarse-to-fine pipeline that captures motions at varying granularities. Moreover, a rate-distortion optimization is adopted to achieve an optimal balance between bitrate consumption and representation fidelity. Experimental results demonstrate that ADC-GS outperforms the per-Gaussian deformation approaches in rendering speed by 300%-800% while achieving state-of-the-art storage efficiency without compromising rendering quality. The code is released at https://github.com/H-Huang774/ADC-GS.git.

Figures (12)

• Figure 1: Comparison with concurrent dynamic scene reconstruction methods on the HyperNeRF dataset. Our method achieves the smallest storage size and the highest rendering speed while preserving excellent rendering quality.
• Figure 2: Illustration of local similarities of different features in e-d3dgs. The local similarity is measured by the average cosine distances between a Gaussian primitive and its 20 neighbors with minimal Euclidean distance.
• Figure 3: Overview of our ADC-GS framework. Top: ADC-GS organizes Gaussian primitives into a sparse set of anchors and compact residuals within canonical space. Bottom right: Gaussian primitives used for rendering are deformed from canonical space through a coarse-to-fine strategy based on anchors. Bottom left: Rendering distortion and estimated bitrates from the MEM are jointly minimized to balance rendering quality and storage efficiency.
• Figure 4: Illustration of the proposed MEM for accurate bitrates estimation. AQM refers to the adaptive quantization module.
• Figure 5: Rate-distortion curves of ADC-GS and comparison methods on HyperNeRF. We vary $\lambda_e$ to achieve variable bitrates.