README.md

---
pretty_name: COLMAP Testing Dataset
license: other
task_categories:
  - image-to-3d
  - depth-estimation
tags:
  - structure-from-motion
  - multi-view-stereo
  - colmap
  - sfm
  - mvs
  - 3d-reconstruction
  - photogrammetry
  - eth3d
size_categories:
  - n<1K
---

# COLMAP Testing Dataset

A small, ready-to-run collection of multi-view image scenes for testing and
benchmarking the [COLMAP](https://colmap.github.io/) Structure-from-Motion (SfM)
and Multi-View Stereo (MVS) pipeline — used here to validate a macOS / Apple
Silicon (Metal) build of COLMAP, but useful for any SfM/MVS work.

It bundles two well-known families of scenes, each already laid out in the
directory structure COLMAP expects, so you can point COLMAP at a folder and run
the pipeline end-to-end with no preprocessing.

## What's inside

| Scene | Family | Images | Resolution | Ground truth provided | Size |
|-------|--------|-------:|------------|------------------------|-----:|
| `courtyard` | ETH3D high-res DSLR | 38 | 6048×4032 | calibration + sparse (COLMAP txt) | 414 MB |
| `electro` | ETH3D high-res DSLR | 45 | 6048×4032 | calibration + sparse (COLMAP txt) | 411 MB |
| `kicker` | ETH3D high-res DSLR | 31 | 6048×4032 | calibration + sparse (COLMAP txt) | 394 MB |
| `relief` | ETH3D high-res DSLR | 31 | 6048×4032 | calibration + sparse (COLMAP txt) | 281 MB |
| `terrains` | ETH3D high-res DSLR | 42 | 6048×4032 | calibration + sparse (COLMAP txt) | 350 MB |
| `gerrard-hall` | COLMAP example | 100 | ~3072×2304 | sparse model + prebuilt `database.db` (158 MB) | 1.0 GB |
| `south-building` | COLMAP example | 128 | ~3072×2304 | sparse model + prebuilt `database.db` (211 MB) | 466 MB |

Total: **7 scenes, 415 images, ~3.2 GB.**

## Directory layout

Two layouts, depending on the family:

**ETH3D scenes** (`courtyard`, `electro`, `kicker`, `relief`, `terrains`):

```
<scene>/
├── images/
│   └── dslr_images/           # the input JPGs (6048×4032)
└── dslr_calibration_jpg/      # ground-truth, COLMAP text format
    ├── cameras.txt            # THIN_PRISM_FISHEYE intrinsics (4 cameras)
    ├── images.txt             # GT camera poses (2 lines per image)
    └── points3D.txt           # GT sparse 3D points
```

**COLMAP example scenes** (`gerrard-hall`, `south-building`):

```
<scene>/
├── images/                    # the input JPGs (flat)
├── sparse/                    # reference sparse reconstruction (COLMAP text)
│   ├── cameras.txt
│   ├── images.txt
│   └── points3D.txt
└── database.db                # prebuilt COLMAP database (features + matches)
```

## Quick start with COLMAP

Sparse reconstruction from scratch (works for any scene):

```bash
SCENE=south-building            # or courtyard, electro, ...
IMG=$SCENE/images               # ETH3D: $SCENE/images/dslr_images

colmap feature_extractor --image_path "$IMG" --database_path db.db
colmap exhaustive_matcher  --database_path db.db
colmap mapper --image_path "$IMG" --database_path db.db --output_path sparse
colmap model_analyzer --path sparse/0
```

The COLMAP example scenes ship a prebuilt `database.db` (features already
extracted and matched), so you can skip straight to mapping:

```bash
colmap mapper --image_path south-building/images \
  --database_path south-building/database.db --output_path sparse
```

Dense MVS (after sparse, with a Metal or CUDA build):

```bash
colmap image_undistorter --image_path "$IMG" --input_path sparse/0 --output_path dense
colmap patch_match_stereo --workspace_path dense
colmap stereo_fusion --workspace_path dense --output_path dense/fused.ply
```

## Golden MVS reference (`colmap_golden_bundle/`)

The repo also includes **`colmap_golden_bundle/`** — the CUDA
`patch_match_stereo` **golden output** for the `south-building` scene, used to
validate that a Metal (Apple-GPU) MVS port produces equivalent depth/normal
maps. Generated with upstream COLMAP 4.0.4 (CUDA) on an NVIDIA T4.

```
colmap_golden_bundle/
├── note.md                            # full provenance + how-to (read this first)
├── colmap_cuda_golden_data.ipynb      # the idempotent notebook that produced it
├── golden_mvs/dense/
│   ├── fused.ply                       # 3,609,743 fused points (93 MB)
│   └── stereo/
│       ├── depth_maps/*.geometric.bin  # 128 geometric depth maps (the reference)
│       └── normal_maps/*.geometric.bin # 128 geometric normal maps
└── logs/                               # undistort / patch_match / fusion logs
```

Depth/normal maps are COLMAP dense binary: an ASCII `width&height&channels&`
header followed by **column-major (Fortran-order) float32**, where **0 =
invalid**. To validate a port, run `patch_match_stereo` on `south-building` and
diff your maps against these on overlapping valid pixels, within tolerance. A
ready-to-use Python reader and the full process note live in
`colmap_golden_bundle/note.md`.

## Provenance, licensing & citation

These scenes are **redistributed for testing convenience**; they are not
original to this dataset. Please respect and cite the original sources, and
consult their terms before any non-testing use.

- **ETH3D scenes** (`courtyard`, `electro`, `kicker`, `relief`, `terrains`) come
  from the **ETH3D multi-view stereo benchmark** (high-resolution DSLR subset),
  Schöps *et al.*, *"A Multi-View Stereo Benchmark with High-Resolution Images
  and Multi-Camera Videos"*, CVPR 2017 — <https://www.eth3d.net/>. The
  laser-scan ground-truth surface is **not** included here (only the
  camera calibration and ground-truth sparse points).
- **COLMAP example scenes** (`gerrard-hall`, `south-building`) are the standard
  COLMAP example datasets, Schönberger & Frahm, *"Structure-from-Motion
  Revisited"*, CVPR 2016 — <https://colmap.github.io/datasets.html>.

`license: other` — licensing follows the original sources above, not a single
blanket license.

See **`llm.txt`** for a dense, machine-readable guide to the file formats and
how to load/evaluate against the ground truth.