Reproduce “Differentiable Acoustic Radiance Transfer” on OpenResearch

About this paper

Sungho Lee, Matteo Scerbo, Seungu Han, Min Jun Choi, Kyogu Lee, Enzo De Sena

Differentiable Acoustic Radiance Transfer (DART) turns acoustic radiance transfer into a GPU-trainable, differentiable simulator that learns directional reflection and transmission properties from sparse room measurements. On echogram prediction, it outperforms neural-field, interpolation, and image-source baselines, reaching L1 errors of 0.510 on a random coupled-room split and 0.831 on an unseen-region split, compared with 0.772 and 5.930 for DiffRIR.

Key insights

• Explicit physical structure gives DART a useful inductive bias in sparse-data settings; neural baselines frequently failed to beat simple interpolation, especially for coupled rooms and unseen regions.
• Learning patch-specific directional materials allows the simulator to model scattered energy and partially compensate for moderate geometric errors that fixed material models cannot absorb.
• The unconstrained material model generally performs best, while the parametric model offers greater interpretability and lower flexibility but can impose unrealistic limits on room-specific scattering and transmission.

Summary by alphaXiv.