Inverse Rendering Eval

inverse-rendering-eval

Radiometrically Consistent Gaussian Surfels for Inverse Rendering — Han et al. (2026) (arXiv:2603.01491, 2026)

What this evaluates

Evaluates a model's capability to perform novel view synthesis, decompose scene properties (albedo, normals, roughness), and relight scenes under new lighting conditions using Gaussian surfels. It specifically probes the model's ability to model indirect illumination and inter-reflections without relying on pre-trained novel view synthesis data.

Datasets

• TensoIR* — total ?; splits: test (-1)
• Synthetic4Relight* — total ?; splits: test (-1)

Metrics

• PSNR (primary) — range: [0, ∞)
  • Peak Signal-to-Noise Ratio computed in dB between predicted and ground truth images. Higher values indicate better reconstruction quality.
• SSIM — range: [0, 1]
  • Structural Similarity Index measuring perceived image quality based on luminance, contrast, and structure. Values range from 0 to 1, with 1 being identical.
• LPIPS — range: [0, 1]
  • Learned Perceptual Image Patch Similarity using deep network features to measure perceptual distance. Lower values indicate higher perceptual similarity.
• MAE — range: [0, 90]
  • Mean Angular Error between predicted and ground truth surface normals, measured in degrees. Lower values indicate better normal estimation.
• MSE — range: [0, ∞)
  • Mean Square Error between predicted and ground truth roughness maps. Lower values indicate better roughness estimation.

Input / output format

Input: Multi-view RGB images with known camera poses and lighting conditions for training; novel viewpoints and target lighting directions for evaluation.

Output: Rendered novel-view images, decomposed material maps (albedo, roughness, normals), and relit images under specified new lighting conditions.

Scoring recipe

def compute_metrics(pred_img, gt_img, pred_normal=None, gt_normal=None, pred_rough=None, gt_rough=None):
    psnr = 10 * log10(1.0 / mse(pred_img, gt_img))
    ssim = structural_similarity(pred_img, gt_img, data_range=1.0)
    lpips = perceptual_loss(pred_img, gt_img)
    mae = mean_angle_error(pred_normal, gt_normal) if pred_normal else None
    mse_rough = mean_squared_error(pred_rough, gt_rough) if pred_rough else None
    return {'PSNR': psnr, 'SSIM': ssim, 'LPIPS': lpips, 'MAE': mae, 'MSE': mse_rough}

Common pitfalls

• Removing the radiometric consistency loss degrades performance on unobserved views and indirect illumination, particularly hurting albedo reconstruction.
• Finetuning-based relighting trades off visual fidelity for speed by accumulating geometric/material estimation errors into surfel radiances, which can mislead readers comparing it to ray-tracing based relighting.

Evidence (verbatim from paper)

We employ PSNR, SSIM, and LPIPS for evaluating NVS, albedo, and relighting. Normal reconstruction is evaluated using Mean Angular Error (MAE), and roughness is evaluated using Mean Square Error (MSE).

Citation

@misc{han2026radiometric,
  title={Radiometrically Consistent Gaussian Surfels for Inverse Rendering},
  author={Han et al. (2026)},
  year={2026},
  note={arXiv:2603.01491}
}

• arXiv: 2603.01491