Human-in-the-Loop Gaussian Model Enhancement with Mobile Robotic Re-Capture

3D Gaussian Splatting can create virtual models with captured real-world images using a camera mounted on a mobile robotic arm. However, the Gaussian model quality deteriorates when parts of the object that have not been previously captured are exposed in VR interaction. This poster presents a human-in-the-loop Gaussian model enhancement method, comprising pre-training, viewpoint labeling, robotic image re-capture, and Gaussian model enhancement. By incorporating newly re-captured images, the enhanced models achieve improved image similarity metrics compared to the pre-trained ones, with minimal optimization time. This method facilitates continuous quality improvement, making Gaussian models more adaptable for VR applications.