**Gaussian in the Wild (GS-W): 3D Gaussian Splatting for Unconstrained Image Collections** **Authors:** Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang **Institution:** Tsinghua Shenzhen International Graduate School, Tsinghua University **Abstract:** Novel view synthesis from unconstrained in-the-wild images remains a challenging task due to photometric variations and transient occluders. Previous methods often introduce a global appearance feature in Neural Radiance Fields (NeRF) to handle appearance variations, but this approach struggles with local high-frequency changes and dynamic environmental factors. Inspired by the unique appearance of each point in a scene, GS-W uses 3D Gaussian points to reconstruct the scene, introducing separated intrinsic and dynamic appearance features for each point. An adaptive sampling strategy allows each Gaussian point to focus on local and detailed information, and a 2D visibility map reduces the impact of transient objects. Experimental results demonstrate that GS-W outperforms NeRF-based methods in terms of reconstruction quality and rendering speed. **Contributions:** - Proposes a new framework GS-W, a 3D Gaussian Splatting-based method, where each Gaussian point has separated intrinsic and dynamic appearance features. - Introduces an adaptive sampling strategy to allow each point to focus on diverse detailed dynamic appearance information. - Reduces the impact of transient objects using a 2D visibility map. - Achieves better reconstruction quality and details compared to NeRF-based methods with a faster rendering speed. **Keywords:** Novel view synthesis, 3D Gaussian Splatting, Unconstrained image collections **Introduction:** Novel view synthesis from unconstrained images is a challenging task due to the dynamic and complex nature of real-world scenes. While NeRF-based methods have shown impressive progress, they often struggle with local details and high-frequency changes. GS-W addresses these issues by using 3D Gaussian points and separating intrinsic and dynamic appearance features, along with an adaptive sampling strategy and a visibility map to handle transient objects. **Related Work:** The paper discusses the limitations of previous methods, such as NeRF-based approaches, and highlights the need for more flexible and detailed appearance modeling. It also reviews 3D representations and novel view synthesis techniques, emphasizing the importance of handling appearance variations and transient objects. **Preliminaries:** The paper introduces 3D Gaussian Splatting (3DGS) and explains how it represents scenes using explicit 3D Gaussian points. It details the projection feature map and K feature maps used to extract high-dimensional sampling space, enabling adaptive sampling and dynamic appearance feature modeling. **Method:** GS-W uses 3D Gaussian points to model the scene, introduces separated intrinsic and dynamic appearance features, and proposes an adaptive sampling strategy to capture local environmental factors. A 2D visibility map is used to handle transient objects.**Gaussian in the Wild (GS-W): 3D Gaussian Splatting for Unconstrained Image Collections** **Authors:** Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang **Institution:** Tsinghua Shenzhen International Graduate School, Tsinghua University **Abstract:** Novel view synthesis from unconstrained in-the-wild images remains a challenging task due to photometric variations and transient occluders. Previous methods often introduce a global appearance feature in Neural Radiance Fields (NeRF) to handle appearance variations, but this approach struggles with local high-frequency changes and dynamic environmental factors. Inspired by the unique appearance of each point in a scene, GS-W uses 3D Gaussian points to reconstruct the scene, introducing separated intrinsic and dynamic appearance features for each point. An adaptive sampling strategy allows each Gaussian point to focus on local and detailed information, and a 2D visibility map reduces the impact of transient objects. Experimental results demonstrate that GS-W outperforms NeRF-based methods in terms of reconstruction quality and rendering speed. **Contributions:** - Proposes a new framework GS-W, a 3D Gaussian Splatting-based method, where each Gaussian point has separated intrinsic and dynamic appearance features. - Introduces an adaptive sampling strategy to allow each point to focus on diverse detailed dynamic appearance information. - Reduces the impact of transient objects using a 2D visibility map. - Achieves better reconstruction quality and details compared to NeRF-based methods with a faster rendering speed. **Keywords:** Novel view synthesis, 3D Gaussian Splatting, Unconstrained image collections **Introduction:** Novel view synthesis from unconstrained images is a challenging task due to the dynamic and complex nature of real-world scenes. While NeRF-based methods have shown impressive progress, they often struggle with local details and high-frequency changes. GS-W addresses these issues by using 3D Gaussian points and separating intrinsic and dynamic appearance features, along with an adaptive sampling strategy and a visibility map to handle transient objects. **Related Work:** The paper discusses the limitations of previous methods, such as NeRF-based approaches, and highlights the need for more flexible and detailed appearance modeling. It also reviews 3D representations and novel view synthesis techniques, emphasizing the importance of handling appearance variations and transient objects. **Preliminaries:** The paper introduces 3D Gaussian Splatting (3DGS) and explains how it represents scenes using explicit 3D Gaussian points. It details the projection feature map and K feature maps used to extract high-dimensional sampling space, enabling adaptive sampling and dynamic appearance feature modeling. **Method:** GS-W uses 3D Gaussian points to model the scene, introduces separated intrinsic and dynamic appearance features, and proposes an adaptive sampling strategy to capture local environmental factors. A 2D visibility map is used to handle transient objects.