Contranerf Eval

contranerf-eval

ContraNeRF: Generalizable Neural Radiance Fields for Synthetic-to-real Novel View Synthesis via Contrastive Learning — Hao Yang et al. (arXiv:2303.11052, 2023)

What this evaluates

Evaluates generalizable Neural Radiance Field (NeRF) methods for novel view synthesis, specifically probing their ability to generalize from synthetic training data to real-world indoor and outdoor scenes. It measures rendering quality and geometric consistency across different domain gaps.

Datasets

• 3D-FRONT — total 88; splits: train (88)
• ScanNet — total 96; splits: train (88), test (8)
• DTU — total 4; splits: test (4)
• LLFF — total ?; splits: test (-1)
• Google Scanned Object — total ?; splits: train (-1)

Metrics

• PSNR (primary) — range: dB / [0, 1]
  • Peak Signal-to-Noise Ratio, measuring pixel-wise reconstruction accuracy in decibels. Higher values indicate better fidelity.
• SSIM — range: [0, 1]
  • Structural Similarity Index, measuring perceptual structural similarity between rendered and ground truth images. Ranges from 0 to 1, where 1 is perfect match.
• LPIPS — range: [0, 1]
  • Learned Perceptual Image Patch Similarity, measuring perceptual distance using deep features. Lower values indicate higher perceptual similarity.

Input / output format

Input: A set of source RGB images with known camera poses for novel view synthesis. For ScanNet, 10 neighboring input views are used. For DTU/LLFF, standard multi-view inputs are provided.

Output: Rendered RGB image of the novel view at specified resolution (640×480 for ScanNet, 800×600 for DTU, 1008×756 for LLFF).

Scoring recipe

def compute_metrics(rendered_img, gt_img):
    psnr = peak_signal_noise_ratio(gt_img, rendered_img)
    ssim = structural_similarity(gt_img, rendered_img, data_range=gt_img.max() - gt_img.min())
    lpips = lpips_loss(rendered_img, gt_img)
    return psnr, ssim, lpips

Common pitfalls

• Synthetic-trained models produce sharper but geometrically inaccurate volume densities due to domain gap.
• Interpolation-based methods may artificially inflate color prediction performance because they directly copy input colors, masking domain shift.
• Fair comparison requires running all baselines on identical experiment settings using official code.

Evidence (verbatim from paper)

Following IBRNet, we evaluate all these methods using PSNR, SSIM, and LPIPS. We report PSNR/SSIM (higher is better) and LPIPS (lower is better).

Citation

@misc{yang2023contranerf,
  title={ContraNeRF: Generalizable Neural Radiance Fields for Synthetic-to-real Novel View Synthesis via Contrastive Learning},
  author={Hao Yang et al.},
  year={2023},
  note={arXiv:2303.11052}
}

• arXiv: 2303.11052