Efficient inverse volume rendering via path filtering for optical property estimation in scattering media

We present a differentiable inverse volume rendering framework for estimating optical properties of scattering media with micron-scale particles (1-3 μm diameter) by integrating tabulated Lorenz-Mie phase functions into a Monte Carlo renderer. To improve accuracy and efficiency, we introduce path filtering-grouping photon paths by distance, scatter count, and energy, and retaining the direct and mid-length classes while discarding low-information deep-scattering paths. In our tested capture geometry and optical regimes, seven views (0° ± 15° ± 30° ± 45°) are sufficient for stable estimation. We report quantitative results with sub-5% normalized error on NIST-traceable microspheres and appearance-consistent reconstructions for representative everyday turbid liquids, with optimization completing within minutes on a single GPU for our path-filtered variant; unfiltered baselines may take longer.