Evidential Nerf Eval

evidential-nerf-eval

Evidential Neural Radiance Fields — Duan et al. (2026) (arXiv:2602.23574, 2026)

What this evaluates

Evaluates the ability of neural radiance field (NeRF) models to accurately reconstruct 3D scenes from 2D images while simultaneously quantifying both aleatoric (data noise) and epistemic (model ignorance) uncertainties. It probes whether uncertainty estimates reliably correlate with actual rendering errors and calibration across varying scene conditions and data sparsity.

Datasets

• Light Field (LF) — total ?; splits: train (-1), test (-1)
• Local Light Field Fusion (LLFF) — total ?; splits: train (-1), test (-1)
• RobustNeRF — total ?; splits: train (-1), test (-1)

Metrics

• PSNR — range: other
  • Peak Signal-to-Noise Ratio computed on rendered vs. ground truth images. Higher is better.
• SSIM — range: [0, 1]
  • Structural Similarity Index measuring perceptual image quality. Range [0, 1].
• LPIPS — range: [0, 1]
  • Learned Perceptual Image Patch Similarity using deep features. Lower is better.
• NLL (primary) — range: other
  • Negative Log-Likelihood: -log p(y|x) assuming a Gaussian predictive distribution. Measures distributional fit; lower is better.
• AUSE — range: [0, 1]
  • Area Under Sparsification Error curve, computed with respect to both RMSE and MAE. Measures error ranking quality; lower is better.
• AUCE — range: [0, 1]
  • Area Under Calibration Error curve. Measures calibration accuracy; lower is better.

Input / output format

Input: A set of 2D training images (views) capturing a 3D scene, plus a set of held-out test views for evaluation.

Output: A rendered 2D image from a novel test view, accompanied by per-pixel aleatoric and epistemic uncertainty maps.

Scoring recipe

def compute_metrics(pred_img, pred_mean, pred_var, gt_img):
    psnr = 10 * log10(1.0 / mse(pred_img, gt_img))
    ssim = structural_similarity(pred_img, gt_img)
    lpips = lpips_loss(pred_img, gt_img)
    errors = (pred_img - gt_img) ** 2
    nll = mean(0.5 * log(2 * pi * pred_var) + 0.5 * errors / pred_var)
    ause_rmse = area_under_sparsification(errors, sqrt(pred_var), metric='rmse')
    ause_mae = area_under_sparsification(errors, sqrt(pred_var), metric='mae')
    auce = area_under_calibration_error(pred_mean, pred_var, gt_img)
    return psnr, ssim, lpips, nll, ause_rmse, ause_mae, auce

Common pitfalls

• Prior benchmarks use varying data splits, architectures, and training schemes, making direct comparison difficult without a standardized setup.
• Ensemble methods often achieve the best UQ metrics but are computationally prohibitive, skewing efficiency comparisons if not accounted for.
• Dropout implementations may lack hardware acceleration (e.g., Tiny CUDA Neural Networks), artificially inflating training/inference times compared to other baselines.

Evidence (verbatim from paper)

For images, we report PSNR, SSIM, and LPIPS to reflect the image reconstruction quality. For uncertainties, we use negative log-likelihood (NLL), area under sparsification error (AUSE) with respect to both RMSE and MAE, and area under calibration error (AUCE), measuring the quality of uncertainty estimates in terms of distributional fit, error ranking, and calibration accuracy.

Citation

@misc{duan2026evidentialnerf,
  title={Evidential Neural Radiance Fields},
  author={Duan et al. (2026)},
  year={2026},
  note={arXiv:2602.23574}
}

• arXiv: 2602.23574