Efficient 3DGS object segmentation via COLMAP point cloud

3D Gaussian Splatting (3DGS), as an emerging 3D scene representation method, has gained significant attention for its high-quality rendering effects and real-time rendering speed. Object segmentation in 3DGS scenes plays a crucial role in scene understanding and editing. For single-object scenes where objects are placed on a plane and captured in 360 degrees, we propose a novel object segmentation approach that directly leverages geometric features in 3D space through plane constraints. Our key insight is to utilize COLMAP sparse point clouds instead of optimized 3DGS points, as they are less affected by environmental factors (such as lighting variations across different views caused by auto-exposure adjustments, background reflections, etc.) in scenes directly captured by smartphones, thus providing more reliable geometric information. With just three user-selected plane points in a single view, our method can determine the reference plane in COLMAP point clouds, extract the target object point cloud region through clustering, and complete the segmentation by extracting 3DGS points within the obtained 3D contour. Experiments demonstrate that our method effectively segments objects without requiring 2D annotations or additional training, achieving excellent segmentation results on both public datasets and smartphone-captured scenes, facilitating the understanding and editing of objects in 3D environments.