VisibilitySplat
An occlusion modeling method for 3D Gaussian splatting inverse rendering. It introduces a sparse BVH-based ray tracing visibility oracle to handle occlusion queries, supporting decoupling of materials and lighting as well as relighting.
Timeline (last 90 days)
VisibilitySplat proposes a method that retains the efficient training and visualization pipeline of rasterized 3DGS while additionally attaching a sparse visibility "oracle" to handle occlusion queries.
The core component of VisibilitySplat is a point-based ray tracer, accelerated with a BVH, used to estimate the transmittance-weighted occlusion of Gaussian primitives under sampled lighting directions.
VisibilitySplat is compared against multiple classes of baselines on the TensoIR synthetic dataset, achieving competitive performance on the three tasks of novel view synthesis, albedo reconstruction, and relighting, and outperforming several earlier baselines on multiple metrics.
Ablation experiments show that explicit visibility modeling makes a substantial contribution to relighting and material-light disentanglement, and that removing it causes noticeable degradation on these two tasks.
Sparse querying means occlusion accuracy is lossy, and especially in scenes with complex geometry and dense mutual occlusion, there is room for trade-offs between sampling density and final disentanglement quality.
All Facts (5)
VisibilitySplat proposes a method that retains the efficient training and visualization pipeline of rasterized 3DGS while additionally attaching a sparse visibility "oracle" to handle occlusion queries.
50%UnverifiedThe core component of VisibilitySplat is a point-based ray tracer, accelerated with a BVH, used to estimate the transmittance-weighted occlusion of Gaussian primitives under sampled lighting directions.
50%UnverifiedVisibilitySplat is compared against multiple classes of baselines on the TensoIR synthetic dataset, achieving competitive performance on the three tasks of novel view synthesis, albedo reconstruction, and relighting, and outperforming several earlier baselines on multiple metrics.
50%UnverifiedAblation experiments show that explicit visibility modeling makes a substantial contribution to relighting and material-light disentanglement, and that removing it causes noticeable degradation on these two tasks.
50%UnverifiedSparse querying means occlusion accuracy is lossy, and especially in scenes with complex geometry and dense mutual occlusion, there is room for trade-offs between sampling density and final disentanglement quality.
50%