A Preliminary Study: Gaussian Splatting Technique in Generating a 3D Model of Textureless Object

Abstract

The traditional static method faces limitations and challenges in reconstructing a 3D model of a real object with a textureless surface, in various environmental lighting conditions. Gaussian Splatting is approached as a superior method to reduce the processing time and draw back those limitations. To assess the visualization and accuracy, a project was designed to create a 3D model of a wooden board and a metal cube under different lighting scenarios using Gaussian Splat Creator & Viewer web-based software. The results show that the model can obviously represent the object’s color, shape, and shadow spot. The additional light source can increase the brightness, and draw different impacts on the models to recognize the fine features on the surface. Furthermore, the translucent effect was exposed on all models, except the wooden board without an additional light source. The models have RMSE less than 6 cm and are categorized as the Level of Detail 4. The highest accuracy was achieved by performing the adjusting scale option, even though a rescaling factor is not defined. Meanwhile, defining the rescale factor and neglecting the scale adjustment process simultaneously, can decrease the accuracy. Last, the 3D reconstruction process of 300 to 400 images takes around 30 to 45 minutes, and the storage size is 1 to 3 Megabytes only.