Spherical Voronoi Radiance Eval

spherical-voronoi-radiance-eval

Spherical Voronoi: Directional Appearance as a Differentiable Partition of the Sphere — Francesco Di Sario et al. (2025) (arXiv:2512.14180, 2025)

What this evaluates

This evaluation probes a differentiable spherical Voronoi partition for modeling view-dependent appearance and reflections in 3D Gaussian Splatting. It measures novel-view synthesis reconstruction fidelity and rendering efficiency against established radiance field baselines across synthetic and real-world scenes.

Datasets

• Mip-NeRF360 — total ?; splits: test (-1)
• DeepBlending — total ?; splits: test (-1)
• Tanks&Temples — total ?; splits: test (-1)
• NeRF-Synthetic — total ?; splits: test (-1)
• Ref-NeRF — total ?; splits: test (-1)
• GlossySynthetic — total ?; splits: test (-1)
• Ref-Real — total ?; splits: test (-1)

Metrics

• PSNR (primary) — range: [0, 100]
  • Peak Signal-to-Noise Ratio in decibels, computed as 10 * log10(MAX^2 / MSE) between rendered and ground truth images.
• SSIM — range: [0, 1]
  • Structural Similarity Index measuring perceived structural changes between two images, averaged over local windows.
• LPIPS — range: [0, 1]
  • Learned Perceptual Image Patch Similarity using deep features to measure perceptual distance between images.

Input / output format

Input: Training images with corresponding camera poses and 3D Gaussian parameters (position, covariance, opacity, color, view-dependent coefficients) used to render novel viewpoints.

Output: Rendered RGB images from specified novel camera viewpoints for direct pixel-wise and perceptual comparison against ground truth.

Scoring recipe

import torch
from torchvision.transforms.functional import psnr, ssim
from lpips import LPIPS

def compute_metrics(rendered, ground_truth):
    rendered = rendered.clamp(0, 1)
    ground_truth = ground_truth.clamp(0, 1)
    p = psnr(rendered, ground_truth, data_range=1.0)
    s = ssim(rendered, ground_truth, data_range=1.0)
    l = lpips_model(rendered, ground_truth)
    return p, s, l

Common pitfalls

• Baselines use test-set feedback for early stopping (every 500 iterations), while the proposed method uses fixed iteration training without test-set feedback, creating an asymmetric evaluation protocol.
• Standard photometric and perceptual metrics (PSNR, SSIM, LPIPS) may not fully capture qualitative improvements in high-frequency specular reflections and local light interactions.
• Memory and parameter counts are reported per Gaussian, but total scene memory scales with the number of Gaussians, which varies across datasets.

Evidence (verbatim from paper)

Our Spherical Voronoi formulation yields consistent improvements over all baselines, achieving higher PSNR than all other color parameterizations (SH, SG, SB).

Citation

@misc{disario2025sphericalvoronoi,
  title={Spherical Voronoi: Directional Appearance as a Differentiable Partition of the Sphere},
  author={Francesco Di Sario et al. (2025)},
  year={2025},
  note={arXiv:2512.14180}
}

• arXiv: 2512.14180