Autonomous Image-based Scanning Pipeline for 3d Digitization of Transparent Objects Using Scene Representation in the Cultural Heritage Domain

3d reconstructions are crucial in various fields such as cultural heritage digitization, but also in medicine, and industry for damage assessment, robotic interaction, and precise documentation. However, the reconstruction of transparent objects is challenging for conventional image-based methods like Multi-View Stereo due to light refraction, limiting feature-based technologies. To address these challenges, this paper introduces an autonomous scanning pipeline for accurately digitizing transparent objects, utilizing direct camera pose access during image acquisition through precise hand-eye calibration of a mechatronic system. This enables clean neural surface and Gaussian Splat reconstructions with implicit geometry estimation. The paper explores input datasets under varying lighting conditions and employs various 3d reconstruction methods to evaluate transparent 3d printed objects and vintage drinking glasses. We demonstrate that the proposed scanning process, combined with optimal lighting, enables accurate training and extraction of 3d reconstructions for transparent objects, without the need for specialized neural surface network structures.