“…In particular, we do not use primary or shadow rays because they are too view-dependent. We use 22 test scenes and 9 tree building algorithms: fast spatial mean-based LBVH [Lauterbach et al 2009;Karras 2012], two greedy sweep-based top-down algorithms (BBVH) [MacDonald and Booth 1990] and SBVH [Stich et al 2009], one bottom-up algorithm (Agglo) [Walter et al 2008], postprocess improvement using iterative reinsertion (Bittner) [Bittner et al 2013], two methods based on local tree rotations: hill climbing and simulated annealing [Kensler 2008], and treelet restructuring [Karras and Aila 2013] with and without triangle splitting (STreelet and Treelet, respectively). The last five are initialized using LBVH, and all algorithms use our own implementations relying on the default parameters recommended by the authors.…”