ICLR 2027 Submission #53498 · Anonymous Rebuttal Materials
Reviewers: SseVouDjncs9
Video 1. [SseV W3] Comparison with Ref-DGS and MaterialRefGS on the office scene.
Video 2. [SseV W2, W3] Comparison with Ref-DGS and MaterialRefGS on the outdoor parking scene (Audi).
Fig. 1. [SseV W2] Qualitative comparison on new real-world captures (Desk and Cabinet).
Hover over the zoomed region to compare.
Fig. 2. [SseV W4, Q1; ncs9 W3, Q3] Comparison with Ref-DGS, and uncertainty maps, selected prompts, and masks of UGMS on real dielectric objects (Helmet and TV Screen).
Hover over the zoomed region to compare.
Video 3. [ncs9 W1, Q1] Comparison of PCC and scene-based ray tracing on reflections (Vase). Ref-Gaussian and MaterialRefGS trace secondary rays through the reconstructed Gaussian scene to obtain indirect light (bottom row). The traced indirect light is noisy and broken, so the reflections on the vase are blurred, whereas ours keeps them sharp without ray tracing.
Video 4. [ncs9 W2, Q2] Ours vs. scene-based ray tracing on the Vase scene. Left: slider comparison of the final renderings (ours on the left, ray tracing on the right). Right: the indirect light obtained by ray tracing through the reconstructed Gaussian scene. Its artifacts carry over to the reflections on the vase, while ours reflects the surrounding room cleanly.
Fig. 3. [ncs9 W3, Q3] Uncertainty maps and UGMS masks under training-data perturbations (IndoorRef12: Knight and Rabbit; real captures: Desk and Helmet360). IoU is for the shown view.