Reproduce “ESVR: 3D Ellipsoid-based Sparse Volume Rendering via Structure-aware Primitive Learning and Per-primitive Ray Sampling” on OpenResearch

About this paper

Suemin Jeon, Youjin Kim, Jungwoo Park, Kyungryun Lee, Won-Ki Jeong

ESVR introduces an ellipsoid-based framework for 3D sparse volume rendering, directly learning volumetric intensity fields and enabling accurate, interactive transfer function control via structure-aware primitive learning and per-primitive ray sampling. It achieves high-fidelity reconstruction with up to four orders of magnitude compression and real-time rendering performance for gigabyte-scale datasets while maintaining a low memory footprint.

Key insights

• Directly modeling the 3D volumetric intensity field using primitives, rather than image-level supervision, is essential for preserving full information and enabling accurate, interactive transfer function control.
• Differentiable ellipsoidal primitives, combined with structure-aware importance scoring (including fractional anisotropy) for adaptive density control, lead to highly compact and structurally faithful representations of sparse volumes.
• Accurate volumetric rendering requires per-primitive ray sampling and 3D transfer function evaluation, as 2D rasterization and pre-classification in screen space introduce artifacts and limit flexibility.

Summary by alphaXiv.