Abstract
Simultaneous relighting and novel-view rendering of digital humanrepresentations is an important yet challenging task with numerousapplications. Progress in this area has been significantly limited due to thelack of publicly available, high-quality datasets, especially for full-bodyhuman captures. To address this critical gap, we introduce the HumanOLATdataset, the first publicly accessible large-scale dataset of multi-viewOne-Light-at-a-Time (OLAT) captures of full-body humans. The dataset includesHDR RGB frames under various illuminations, such as white light, environmentmaps, color gradients and fine-grained OLAT illuminations. Our evaluations ofstate-of-the-art relighting and novel-view synthesis methods underscore boththe dataset's value and the significant challenges still present in modelingcomplex human-centric appearance and lighting interactions. We believeHumanOLAT will significantly facilitate future research, enabling rigorousbenchmarking and advancements in both general and human-specific relighting andrendering techniques.