Cvt Xrf Eval

---
name: cvt-xrf-eval
description: Evaluates the quality of novel view synthesis from sparse input views using 3D radiance fields. It measures how well the model reconstructs unseen images and maintains 3D consistency across different sparsity levels (3, 6, or 9 input views). Use when the user wants to benchmark on DTU dataset, Synthetic dataset, or asks about evaluating this task. Reports PSNR.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2403.16885
  bibtex_key: zhong2024cvt_xrf
  confidence: high
---

# cvt-xrf-eval

> CVT-xRF: Contrastive In-Voxel Transformer for 3D Consistent Radiance Fields from Sparse Inputs — Zhong et al. (2024) (arXiv:2403.16885, 2024)

## What this evaluates

Evaluates the quality of novel view synthesis from sparse input views using 3D radiance fields. It measures how well the model reconstructs unseen images and maintains 3D consistency across different sparsity levels (3, 6, or 9 input views).

## Datasets

- **DTU dataset** — total ?; splits: test (-1)
- **Synthetic dataset** — total ?; splits: test (-1)

## Metrics

- `PSNR` **(primary)** — range: other
  - Mean Peak Signal-to-Noise Ratio over all scenes and test views. Measures pixel-wise reconstruction fidelity in decibels (dB).
- `SSIM` — range: [0, 1]
  - Mean Structural Similarity Index over all scenes and test views. Measures perceptual structural similarity between rendered and ground truth images, ranging from 0 to 1.

## Input / output format

**Input**: Sparse set of input views (3, 6, or 9 images) with corresponding camera poses for a target scene.

**Output**: Rendered novel view images (full-image and foreground object masks).

## Scoring recipe

```python
def evaluate(rendered_imgs, gt_imgs):
    psnr_scores = [compute_psnr(r, g) for r, g in zip(rendered_imgs, gt_imgs)]
    ssim_scores = [compute_ssim(r, g) for r, g in zip(rendered_imgs, gt_imgs)]
    return mean(psnr_scores), mean(ssim_scores)
```

## Common pitfalls

- Metrics are averaged across all scenes rather than reported per-scene, which can mask scene-specific performance variations.
- Evaluation is strictly stratified by the number of input views (3, 6, 9 for DTU; 3, 8 for Synthetic), so results are not directly comparable across different sparsity settings.
- Both full-image and foreground object metrics are reported separately; citing only one without specifying the mask type can misrepresent performance.

## Evidence (verbatim from paper)

> For quantitative comparison of synthesis results, we report the mean of peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) [47] over different scenes.

## Citation

```bibtex
@misc{zhong2024cvt_xrf,
  title={CVT-xRF: Contrastive In-Voxel Transformer for 3D Consistent Radiance Fields from Sparse Inputs},
  author={Zhong et al. (2024)},
  year={2024},
  note={arXiv:2403.16885}
}
```

- arXiv: 2403.16885