Abstract
Volumetric rendering has become central to modern novel view synthesismethods, which use differentiable rendering to optimize 3D scenerepresentations directly from observed views. While many recent works build onNeRF or 3D Gaussians, we explore an alternative volumetric scenerepresentation. More specifically, we introduce two new scene representationsbased on linear primitives-octahedra and tetrahedra-both of which definehomogeneous volumes bounded by triangular faces. This formulation alignsnaturally with standard mesh-based tools, minimizing overhead for downstreamapplications. To optimize these primitives, we present a differentiablerasterizer that runs efficiently on GPUs, allowing end-to-end gradient-basedoptimization while maintaining realtime rendering capabilities. Throughexperiments on real-world datasets, we demonstrate comparable performance tostate-of-the-art volumetric methods while requiring fewer primitives to achievesimilar reconstruction fidelity. Our findings provide insights into thegeometry of volumetric rendering and suggest that adopting explicit polyhedracan expand the design space of scene representations.