3D Gaussian object segmentation via COLMAP point cloud classification

Object segmentation in 3D scenes remains an important research topic in computer vision. With the emergence of 3D Gaussian Splatting (3DGS) in scene reconstruction, object segmentation methods based on 3DGS have gained increasing attention. This paper proposes a novel segmentation paradigm that utilizes the initial COLMAP point cloud from the 3DGS reconstruction process to achieve scene segmentation. Specifically, we first embed semantic features into the COLMAP point cloud, train these features and a classifier under the guidance of 2D masks, segment the target object’s corresponding COLMAP point cloud using the trained classifier to obtain its 3D contour, and finally guide the object segmentation in 3DGS based on this contour. This method explores the possibility of utilizing COLMAP point cloud prior to 3DGS reconstruction to assist final 3DGS segmentation, providing a new perspective for 3DGS scene understanding. Experiments demonstrate that our method can effectively accomplish object segmentation tasks in 3DGS scenes, validating the feasibility of this segmentation paradigm.