Nvs Ho Eval

---
name: nvs-ho-eval
description: Evaluates novel view synthesis (NVS) methods on real-world handheld objects using only RGB inputs. It probes a model's ability to reconstruct 3D-consistent renderings from unconstrained, handheld camera trajectories that exhibit motion blur, occlusions, and pose estimation inaccuracies. Use when the user wants to benchmark on NVS-HO, or asks about evaluating this task. Reports PSNR.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2602.05822
  bibtex_key: ali2026nvs-ho
  confidence: high
---

# nvs-ho-eval

> NVS-HO: A Benchmark for Novel View Synthesis of Handheld Objects — Musawar Ali et al. (arXiv:2602.05822, 2026)

## What this evaluates

Evaluates novel view synthesis (NVS) methods on real-world handheld objects using only RGB inputs. It probes a model's ability to reconstruct 3D-consistent renderings from unconstrained, handheld camera trajectories that exhibit motion blur, occlusions, and pose estimation inaccuracies.

## Datasets

- **NVS-HO** — total ?; splits: train (-1), test (-1)

## Metrics

- `PSNR` **(primary)** — range: other (dB)
  - Peak Signal-to-Noise Ratio computed on masked image pairs (object masks applied to both rendered and ground-truth views). Higher values indicate better pixel-level reconstruction fidelity.
- `SSIM` — range: [0, 1]
  - Structural Similarity Index computed on masked image pairs. Measures perceptual similarity between rendered and ground-truth images, with 1 indicating identical structure.
- `LPIPS` — range: [0, 1]
  - Learned Perceptual Image Patch Similarity computed on masked image pairs. Lower values indicate better perceptual similarity according to deep feature distances.

## Input / output format

**Input**: RGB image sequences of handheld objects, along with camera poses estimated via COLMAP or VGGT.

**Output**: Rendered novel view images for each test viewpoint.

## Scoring recipe

```python
def compute_nvs_metrics(rendered_images, gt_images, object_masks):
    # Align rendered and ground-truth views in a common 3D reference frame
    # Apply object masks to filter background
    rendered_masked = [r * m for r, m in zip(rendered_images, object_masks)]
    gt_masked = [g * m for g, m in zip(gt_images, object_masks)]
    # Compute metrics per image pair
    psnr_scores = [compute_psnr(r, g) for r, g in zip(rendered_masked, gt_masked)]
    ssim_scores = [compute_ssim(r, g) for r, g in zip(rendered_masked, gt_masked)]
    lpips_scores = [compute_lpips(r, g) for r, g in zip(rendered_masked, gt_masked)]
    return mean(psnr_scores), mean(ssim_scores), mean(lpips_scores)
```

## Common pitfalls

- Motion blur and occlusions in handheld sequences significantly degrade pose estimation and rendering quality, leading to low absolute metric values.
- Absolute metric values (e.g., median PSNR ~16) are much lower than in controlled setups, so relative improvements between methods are more meaningful than absolute scores.
- Pose refinement during training yields only marginal gains, indicating that accurate pose initialization is the primary bottleneck rather than optimization strategy.

## Evidence (verbatim from paper)

> The method-agnostic evaluation protocol aligns rendered and ground-truth views in a common 3D reference frame, enabling fair comparison of masked image pairs via PSNR, SSIM, and LPIPS.

## Citation

```bibtex
@misc{ali2026nvs-ho,
  title={NVS-HO: A Benchmark for Novel View Synthesis of Handheld Objects},
  author={Musawar Ali et al.},
  year={2026},
  note={arXiv:2602.05822}
}
```

- arXiv: 2602.05822