Dynamic Nerf Soccer Eval

---
name: dynamic-nerf-soccer-eval
description: Evaluates the ability of dynamic NeRF models to perform photorealistic novel view synthesis in large-scale, dynamic sports environments. It probes spatiotemporal modeling capabilities, specifically how well models handle fast-moving small objects (like a soccer ball) and scale variations across different camera configurations. Use when the user wants to benchmark on Synthetic Soccer Scenes (Multi-Camera), or asks about evaluating this task. Reports PSNR.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2309.06802
  bibtex_key: [REDACTED]
  confidence: high
---

# dynamic-nerf-soccer-eval

> Dynamic NeRFs for Soccer Scenes — Lewin et al. (2023) (arXiv:2309.06802, 2023)

## What this evaluates

Evaluates the ability of dynamic NeRF models to perform photorealistic novel view synthesis in large-scale, dynamic sports environments. It probes spatiotemporal modeling capabilities, specifically how well models handle fast-moving small objects (like a soccer ball) and scale variations across different camera configurations.

## Datasets

- **Synthetic Soccer Scenes (Multi-Camera)** — total ?; splits: train (-1), test (-1)

## Metrics

- `PSNR` **(primary)** — range: dB
  - Peak Signal-to-Noise Ratio computed between the rendered novel view and the ground truth image from a dedicated evaluation camera. Standard logarithmic scale in decibels (dB).

## Input / output format

**Input**: Multi-view images or video frames of synthetic soccer scenes captured by 30 cameras. Training uses rays traced through uniformly sampled pixels or Ray Importance Sampling (IS) to prioritize dynamic content.

**Output**: Rendered novel-view images from a dedicated evaluation camera, typically at 1080p resolution.

## Scoring recipe

```python
def compute_psnr(rendered, gt):
    mse = np.mean((rendered - gt) ** 2)
    if mse == 0:
        return float('inf')
    return 10 * np.log10((255.0 ** 2) / mse)
```

## Common pitfalls

- Using uniform pixel sampling during training causes models to ignore small, fast-moving objects (e.g., the ball), leading to complete reconstruction failure.
- Relying solely on focused metrics (e.g., bounding-box PSNR) can be misleading when object scale varies; global metrics may degrade while visual quality improves.
- Fast motion and scale mismatches in stadium-wide views cause ghosting artifacts and blurring that standard metrics do not fully capture.

## Evidence (verbatim from paper)

> Associated metrics, computed using a dedicated evaluation camera, are reported in Tab. 2. ... While the focused PSNR improves when using importance sampling, the other metrics sometimes degrade, which does not support qualitative results from Fig. 6.

## Citation

```bibtex
@misc{lewin2023dynamicnerfs,
  title={Dynamic NeRFs for Soccer Scenes},
  author={Lewin et al. (2023)},
  year={2023},
  note={arXiv:2309.06802}
}
```

- arXiv: 2309.06802