Frame-wise 3DGS-based Human Point Cloud Generation With Gaussian Primitive-aware Upsampling From Fixed Multiview GoPro Videos

Multiview videos provide useful visual information for reconstructing indoor scenes containing moving human objects as explicit 3D data. In this study, we propose a 3D Gaussian Splatting (3DGS)-based point cloud generation pipeline for a fixed multiview GoPro setup. The proposed pipeline generates frame-wise xyz-red–green–blue (rgb) point cloud sequences that can be used as an input format for downstream point cloud processing systems. Camera parameters are estimated from the first synchronized multiview frame using COLMAP and reused for all frames according to the fixed-camera assumption. A 3DGS representation was optimized for each frame, and an initial point cloud was generated by converting Gaussian center positions and zeroth-order spherical harmonics color coefficients into the xyz-rgb PLY format. However, this Gaussian-center-based conversion does not fully utilize the opacity, scale, and rotation parameters, which can result in sparse human object point clouds. To mitigate this limitation, we introduced a Gaussian primitive-aware point cloud upsampling module that selects reliable Gaussian primitives and generates additional support points along the principal axes of the corresponding Gaussian ellipsoids. The generated point clouds were evaluated using projection-based human-region metrics. Experimental results on an indoor multiview GoPro dataset show that the proposed pipeline generates practical 3D human point cloud sequences and that the Gaussian primitive-aware upsampling module improves projected human-region coverage and rgb reconstruction quality.