Hdr Gopro Eval

---
name: hdr-gopro-eval
description: Evaluates the model's ability to reconstruct high-quality dynamic HDR radiance fields and synthesize novel views and time steps from alternating-exposure monocular videos. It probes exposure-invariant geometric reconstruction, temporal coherence, and radiometric accuracy under extreme exposure variations. Use when the user wants to benchmark on HDR-GoPro, or asks about evaluating this task. Reports PSNR.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2603.08313
  bibtex_key: shin2026hdrnsff
  confidence: high
---

# hdr-gopro-eval

> HDR-NSFF: High Dynamic Range Neural Scene Flow Fields — Shin Dong-Yeon et al. (2026) (arXiv:2603.08313, 2026)

## What this evaluates

Evaluates the model's ability to reconstruct high-quality dynamic HDR radiance fields and synthesize novel views and time steps from alternating-exposure monocular videos. It probes exposure-invariant geometric reconstruction, temporal coherence, and radiometric accuracy under extreme exposure variations.

## Datasets

- **HDR-GoPro** — total ?; splits: train (-1), test (-1)

## Metrics

- `PSNR` **(primary)** — range: other
  - Peak Signal-to-Noise Ratio in dB, calculated as 10 * log10(MAX^2 / MSE), where MAX is the maximum possible pixel value and MSE is the mean squared error between predicted and ground truth images.
- `SSIM` — range: [0, 1]
  - Structural Similarity Index measuring perceived change in structural information, ranging from -1 to 1, with 1 indicating perfect similarity.
- `LPIPS` — range: [0, 1]
  - Learned Perceptual Image Patch Similarity, a perceptual metric using deep network features to measure distance between image patches, where lower values indicate higher perceptual similarity.

## Input / output format

**Input**: Single LDR frame per timestamp and viewpoint for training; held-out viewpoints and odd-indexed frames for testing.

**Output**: Tone-mapped LDR/HDR images at novel viewpoints and interpolated time steps.

## Scoring recipe

```python
def compute_metrics(pred_img, gt_img):
    # pred_img and gt_img are tone-mapped to the same exposure
    mse = np.mean((pred_img - gt_img) ** 2)
    psnr = 10 * np.log10(1.0 / mse)  # assuming normalized [0,1]
    ssim = calculate_ssim(pred_img, gt_img)
    lpips = calculate_lpips(pred_img, gt_img)
    return psnr, ssim, lpips
```

## Common pitfalls

- Tone-mapping must be applied consistently (e.g., Photomatix Pro) to both predictions and ground truth before computing metrics to ensure fair comparison.
- Evaluation requires synthesizing HDR radiance first, then tone-mapping to the target exposure, rather than comparing raw HDR values directly against LDR references.
- Temporal interpolation is evaluated by training only on even-indexed frames and testing on odd-indexed frames, doubling the temporal gap compared to standard frame-by-frame synthesis.

## Evidence (verbatim from paper)

> During testing, the reconstructed HDR radiance is tone-mapped to the target exposure for quantitative comparison (e.g., PSNR, SSIM, LPIPS) against these LDR references.

## Citation

```bibtex
@misc{shin2026hdrnsff,
  title={HDR-NSFF: High Dynamic Range Neural Scene Flow Fields},
  author={Shin Dong-Yeon et al. (2026)},
  year={2026},
  note={arXiv:2603.08313}
}
```

- arXiv: 2603.08313