Dtu Nerf Edge Detection Eval

---
name: dtu-nerf-edge-detection-eval
description: Evaluates the geometric reconstruction quality of Neural Radiance Fields by extracting 3D surfaces or edges using density gradients. It measures how accurately the predicted geometry aligns with ground truth point clouds across diverse real-world objects. Use when the user wants to benchmark on DTU benchmark dataset, or asks about evaluating this task. Reports completeness.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2309.14800
  bibtex_key: [REDACTED]
  confidence: high
---

# dtu-nerf-edge-detection-eval

> 3D Density-Gradient based Edge Detection on Neural Radiance Fields (NeRFs) for Geometric Reconstruction — Jäger et al. (2023) (arXiv:2309.14800, 2023)

## What this evaluates

Evaluates the geometric reconstruction quality of Neural Radiance Fields by extracting 3D surfaces or edges using density gradients. It measures how accurately the predicted geometry aligns with ground truth point clouds across diverse real-world objects.

## Datasets

- **DTU benchmark dataset** — total ?; splits: test (6)

## Metrics

- `completeness` **(primary)** — range: percent
  - Percentage of ground truth points that fall within a specified distance threshold of the predicted surface. Defined in Section 3.4 of the paper.
- `correctness` — range: percent
  - Percentage of predicted surface points that fall within a specified distance threshold of the ground truth point cloud. Defined in Section 3.4 of the paper.

## Input / output format

**Input**: Set of RGB images with corresponding camera poses, used to train and evaluate a NeRF density field.

**Output**: 3D point cloud or surface mesh extracted from the NeRF density field using gradient-based edge detection filters (Sobel, Canny, Laplacian of Gaussian) or global density thresholding.

## Scoring recipe

```python
def compute_completeness(pred_pts, gt_pts, threshold=0.01):
    # Distance from each GT point to nearest predicted point
    dists = cdist(gt_pts, pred_pts)
    return (dists.min(axis=1) < threshold).mean() * 100

def compute_correctness(pred_pts, gt_pts, threshold=0.01):
    # Distance from each predicted point to nearest GT point
    dists = cdist(pred_pts, gt_pts)
    return (dists.min(axis=1) < threshold).mean() * 100
```

## Common pitfalls

- Global density thresholding is highly sensitive to NeRF configuration and requires empirical tuning, often leading to incomplete or noisy surfaces.
- Gradient filter parameters (e.g., Canny thresholds, Gaussian standard deviation) significantly impact the trade-off between surface completeness and correctness.

## Evidence (verbatim from paper)

> We evaluate our framework with first derivative Sobel filter and Canny filter as well as second derivative Laplacian of Gaussian filter against different global density thresholds. Thereby qualitative as well as quantitative results based on completeness and correctness as described in the evaluation Section 3.4 are considered.

## Citation

```bibtex
@misc{jager2023density,
  title={3D Density-Gradient based Edge Detection on Neural Radiance Fields (NeRFs) for Geometric Reconstruction},
  author={Jäger et al. (2023)},
  year={2023},
  note={arXiv:2309.14800}
}
```

- arXiv: 2309.14800