Monerf Eval

monerf-eval

Fast Non-Rigid Radiance Fields from Monocularized Data — Kappel et al. (2022) (arXiv:2212.01368, 2022)

What this evaluates

Evaluates the ability of a neural radiance field model to reconstruct and render novel views of dynamic, non-rigid scenes from monocular video input. It probes spatiotemporal deformation modeling, training efficiency, and perceptual image quality across synthetic and real-world sequences.

Datasets

• D-NeRF — total ?; splits: test (-1)
• MMVA — total ?; splits: test (-1)

Metrics

• PSNR (primary) — range: other
  • Peak signal-to-noise ratio computed per-pixel between rendered novel views and ground truth images. Higher values indicate better reconstruction fidelity.
• SSIM — range: [0, 1]
  • Structural Similarity Index measuring perceived structural changes between images. Higher values indicate better similarity.
• LPIPS — range: [0, 1]
  • Learned Perceptual Image Patch Similarity using deep features to measure perceptual distance. Lower values indicate better perceptual quality.

Input / output format

Input: Monocularized video sequences of dynamic scenes. Synthetic data downsampled to 400×400 pixels; real-world data at full resolution 1028×752 pixels.

Output: Rendered novel-view images at held-out timestamps and camera poses.

Scoring recipe

def compute_metrics(rendered_images, ground_truth_images):
    psnr = 10 * np.log10(255**2 / np.mean((rendered_images - ground_truth_images)**2))
    ssim = structural_similarity(rendered_images, ground_truth_images, data_range=255)
    lpips = perceptual_loss(rendered_images, ground_truth_images)  # VGG-based
    return psnr, ssim, lpips
# Training time and FPS are measured empirically during evaluation.

Common pitfalls

• PSNR and SSIM are higher-is-better, while LPIPS is lower-is-better; misinterpreting the direction leads to incorrect conclusions about model performance.
• Training times for baseline methods are copied from their original papers and were likely measured on different hardware, making direct speed comparisons approximate.
• Synthetic D-NeRF results use half-resolution (400×400) inputs for fair comparison, while real-world MMVA uses full resolution (1028×752), so metrics are not directly comparable across datasets.

Evidence (verbatim from paper)

We report three metrics for quantitative assessment: Peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and learned perceptual image patch similarity (LPIPS). While the PSNR reflects per-pixel the error and is thus closest to the training objective function, SSIM and LPIPS gauge the perceptual reconstruction accuracy from a larger context.

Citation

@misc{kappel2022monerf,
  title={Fast Non-Rigid Radiance Fields from Monocularized Data},
  author={Kappel et al. (2022)},
  year={2022},
  note={arXiv:2212.01368}
}

• arXiv: 2212.01368