**ContextGS: Compact 3D Gaussian Splatting with Anchor Level Context Model** **Authors:** Yufei Wang, Zhihao Li, Lanqing Guo, Wenhan Yang, Alex C. Kot, Bihan Wen **Institution:** Nanyang Technological University, Singapore; PengCheng Laboratory, China **Abstract:** This paper addresses the challenge of efficient compression in 3D Gaussian Splatting (3DGS), a promising framework for novel view synthesis. Existing methods primarily compress neural Gaussians individually, neglecting their spatial dependencies. Inspired by context models used in image compression, the authors propose the first autoregressive model at the anchor level for 3DGS compression. They divide anchors into hierarchical levels, where anchors from coarser levels are used to predict anchors in finer levels, reducing spatial redundancy. Additionally, a low-dimensional quantized feature is introduced as a hyperprior for each anchor to enhance coding efficiency. The proposed method achieves a significant reduction in storage size (up to 15 times) while maintaining or improving rendering quality compared to state-of-the-art methods. **Introduction:** 3DGS represents 3D scenes using neural Gaussians, offering fast rendering speeds and high fidelity. However, the large number of Gaussians and their attributes require effective compression techniques. While previous methods focus on individual compression, ContextGS introduces a context model at the anchor level to leverage spatial dependencies. The method divides anchors into levels, encoding them progressively, and uses a quantized hyperprior feature to further improve efficiency. **Methodology:** - **Anchor Partitioning:** Anchors are divided into levels using a "bottom-up" method, ensuring a traceable mapping relationship between adjacent levels. - **Entropy Coding:** Anchors are encoded using an autoregressive model, where the properties of anchors at finer levels are predicted based on already decoded anchors at coarser levels. - **Hyperprior Feature:** A learnable hyperprior vector is introduced for each anchor to reduce spatial redundancy. - **Training Objective:** The training loss includes the rendering loss and an entropy loss to optimize both quality and size. **Experiments:** - **Implementation Details:** The method is built on Scaffold-GS, with specific hyperparameters set for fair comparison. - **Comparison with Baselines:** ContextGS outperforms existing methods in terms of storage size and rendering quality. - **Ablation Studies:** The effectiveness of the anchor-level context model and hyperprior features is demonstrated through ablation studies. **Conclusion:** ContextGS introduces a novel context model at the anchor level, significantly reducing the storage size of 3DGS models while maintaining or improving rendering quality. The method leverages spatial dependencies and introduces a hyperprior feature to enhance coding efficiency, making 3DGS more accessible and efficient for various applications.**ContextGS: Compact 3D Gaussian Splatting with Anchor Level Context Model** **Authors:** Yufei Wang, Zhihao Li, Lanqing Guo, Wenhan Yang, Alex C. Kot, Bihan Wen **Institution:** Nanyang Technological University, Singapore; PengCheng Laboratory, China **Abstract:** This paper addresses the challenge of efficient compression in 3D Gaussian Splatting (3DGS), a promising framework for novel view synthesis. Existing methods primarily compress neural Gaussians individually, neglecting their spatial dependencies. Inspired by context models used in image compression, the authors propose the first autoregressive model at the anchor level for 3DGS compression. They divide anchors into hierarchical levels, where anchors from coarser levels are used to predict anchors in finer levels, reducing spatial redundancy. Additionally, a low-dimensional quantized feature is introduced as a hyperprior for each anchor to enhance coding efficiency. The proposed method achieves a significant reduction in storage size (up to 15 times) while maintaining or improving rendering quality compared to state-of-the-art methods. **Introduction:** 3DGS represents 3D scenes using neural Gaussians, offering fast rendering speeds and high fidelity. However, the large number of Gaussians and their attributes require effective compression techniques. While previous methods focus on individual compression, ContextGS introduces a context model at the anchor level to leverage spatial dependencies. The method divides anchors into levels, encoding them progressively, and uses a quantized hyperprior feature to further improve efficiency. **Methodology:** - **Anchor Partitioning:** Anchors are divided into levels using a "bottom-up" method, ensuring a traceable mapping relationship between adjacent levels. - **Entropy Coding:** Anchors are encoded using an autoregressive model, where the properties of anchors at finer levels are predicted based on already decoded anchors at coarser levels. - **Hyperprior Feature:** A learnable hyperprior vector is introduced for each anchor to reduce spatial redundancy. - **Training Objective:** The training loss includes the rendering loss and an entropy loss to optimize both quality and size. **Experiments:** - **Implementation Details:** The method is built on Scaffold-GS, with specific hyperparameters set for fair comparison. - **Comparison with Baselines:** ContextGS outperforms existing methods in terms of storage size and rendering quality. - **Ablation Studies:** The effectiveness of the anchor-level context model and hyperprior features is demonstrated through ablation studies. **Conclusion:** ContextGS introduces a novel context model at the anchor level, significantly reducing the storage size of 3DGS models while maintaining or improving rendering quality. The method leverages spatial dependencies and introduces a hyperprior feature to enhance coding efficiency, making 3DGS more accessible and efficient for various applications.