Guest Speaker: Konstantinos Polyzos

Title: Active selection of informative images for efficient Gaussian splatting via black-box optimization
 

Abstract: 3D scene rendering is a fundamental computer vision task with diverse applications, including autonomous driving, robotics, and medical imaging, just to name a few. Gaussian splatting (GS) and its extensions and variants provide outstanding performance in fast 3D scene rendering while meeting reduced storage demands and computational efficiency. While the selection of 2D images capturing the scene of interest is crucial for the proper initialization and training of GS, hence markedly affecting the rendering performance, prior works rely on passive image selection, often leading to redundancy and high computational costs in dense-view settings or insufficient scene coverage in sparse-view scenarios. In the first part of the talk, we will focus on adaptive Bayesian optimization to efficiently optimize black-box and expensive-to-evaluate functions by judiciously adapting to the proper surrogate model as new input-output data are acquired online. Next, we will introduce a novel black-box optimization framework, namely `ActiveInitSplat', that actively selects training images for proper initialization and training of GS. ActiveInitSplat relies on density and occupancy criteria of the resultant 3D scene representation from the selected 2D images, to ensure that the latter are captured from diverse viewpoints, leading to better scene coverage and that the initialized Gaussian functions are well aligned with the actual 3D structure. We will conclude with numerical tests on real-world 3D scenes that showcase the merits of ActiveInitSplat compared to passive GS counterparts in both dense- and sparse-view settings.