# Clatch Sfm Eval

> Evaluates the computational efficiency and 3D reconstruction accuracy of a GPU-accelerated binary feature descriptor (CLATCH) compared to traditional and deep learning-based descriptors within a Structure-from-Motion pipeline. Use when the user wants to benchmark on Photogrammetry Image Sets (8 scenes), or asks about evaluating this task. Reports SfM Scene RMSE (pixels).

- Skill: `qhjqhj00/clatch-sfm-eval` (Agent Skill)
- Install (CLI): `npx skillmds add qhjqhj00/clatch-sfm-eval`
- Raw SKILL.md: https://api.skillmd.com/api/skills/qhjqhj00/clatch-sfm-eval/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: DevOps & Infra
- Author: qhjqhj00 (https://skillmd.com/u/qhjqhj00)
- Updated: 2026-09-08
- Page: https://skillmd.com/skills/qhjqhj00/clatch-sfm-eval

---


# clatch-sfm-eval

> The CUDA LATCH Binary Descriptor: Because Sometimes Faster Means Better — Parker et al. (2016) (arXiv:1609.03986, 2016)

## What this evaluates

Evaluates the computational efficiency and 3D reconstruction accuracy of a GPU-accelerated binary feature descriptor (CLATCH) compared to traditional and deep learning-based descriptors within a Structure-from-Motion pipeline.

## Datasets

- **Photogrammetry Image Sets (8 scenes)** — total ?; splits: test (8)

## Metrics

- `SfM Scene RMSE (pixels)` **(primary)** — range: other
  - Root Mean Square Error of reprojection errors across all matched 3D points in the reconstructed scene, measured in pixels.
- `Descriptor Extraction Time (μs)` — range: other
  - Mean time in microseconds required to extract a single local feature descriptor.
- `Total SfM Pipeline Time (s)` — range: other
  - Total time in seconds for descriptor extraction, matching, and incremental SfM reconstruction.

## Input / output format

**Input**: High-resolution photogrammetry image sets (5,616×3,744 or 3,744×5,616 pixels) for each scene.

**Output**: 3D point cloud and mesh reconstructions, along with per-scene reprojection RMSE and pipeline timing statistics.

## Scoring recipe

```python
def compute_sfm_rmse(reconstructed_3d_points, image_2d_points, camera_params):
    errors = []
    for pt3d, pt2d in zip(reconstructed_3d_points, image_2d_points):
        projected_2d = project_3d_to_2d(pt3d, camera_params)
        errors.append(np.linalg.norm(projected_2d - pt2d))
    return np.sqrt(np.mean(np.square(errors)))
```

## Common pitfalls

- Reprojection RMSE values are very small (~0.5 pixels) and inter-method differences are often <0.1 pixels, making statistical significance difficult to assess on only 8 scenes.
- Pipeline timing is dominated by brute-force nearest neighbor matching, so descriptor extraction speed gains do not linearly translate to total pipeline speedups.
- SIFT uses a different feature detector (DOG) than the other methods (CUDA FAST), which may confound direct speed and accuracy comparisons.

## Evidence (verbatim from paper)

> Table 2 summarizes these results, providing the final scene reprojection RMSE and the total time for descriptor extraction, matching and SfM reconstruction. All these tests were run on our GTX 1080 GPU.

## Citation

```bibtex
@misc{parker2016cudalatch,
  title={The CUDA LATCH Binary Descriptor: Because Sometimes Faster Means Better},
  author={Parker et al. (2016)},
  year={2016},
  note={arXiv:1609.03986}
}
```

- arXiv: 1609.03986