llm.txt

# COLMAP Testing Dataset — usage & data guide (for LLMs/agents and humans)

Purpose: ready-to-run multi-view image scenes for testing/benchmarking the
COLMAP SfM/MVS pipeline. 7 scenes, 415 images, ~3.2 GB, in two families with
two different on-disk layouts. Everything is already in COLMAP's expected
structure — no preprocessing needed.

## Scenes

ETH3D high-res DSLR (images at `<scene>/images/dslr_images/`, 6048x4032 JPG,
ground-truth in `<scene>/dslr_calibration_jpg/`):
- courtyard  (38 images)
- electro    (45 images)
- kicker     (31 images)
- relief     (31 images)
- terrains   (42 images)

COLMAP example scenes (images at `<scene>/images/`, reference model in
`<scene>/sparse/`, prebuilt `<scene>/database.db`):
- gerrard-hall   (100 images, database.db 158 MB)
- south-building (128 images, database.db 211 MB)

## On-disk layout

ETH3D:
  <scene>/images/dslr_images/*.JPG          # inputs (24 MP)
  <scene>/dslr_calibration_jpg/cameras.txt  # GT intrinsics (THIN_PRISM_FISHEYE)
  <scene>/dslr_calibration_jpg/images.txt   # GT poses (2 lines per image)
  <scene>/dslr_calibration_jpg/points3D.txt # GT sparse points

COLMAP examples:
  <scene>/images/*.JPG    # inputs
  <scene>/sparse/{cameras.txt,images.txt,points3D.txt}  # reference sparse model
  <scene>/database.db     # prebuilt SQLite DB: features + matches already done

IMPORTANT path difference: ETH3D images live under `images/dslr_images`, COLMAP
examples under `images`. Pass the correct `--image_path` accordingly.

## COLMAP model file formats (what cameras/images/points3D mean)

These are COLMAP's standard sparse-model text files (same schema everywhere).

cameras.txt — one line per camera:
  CAMERA_ID  MODEL  WIDTH  HEIGHT  PARAMS...
  ETH3D uses MODEL=THIN_PRISM_FISHEYE with 12 params
  (fx fy cx cy k1 k2 p1 p2 k3 k4 sx1 sy1). COLMAP example scenes use simpler
  models (e.g. SIMPLE_RADIAL / PINHOLE) — read the file, don't assume.

images.txt — TWO lines per image (this trips up naive parsers):
  line 1:  IMAGE_ID  QW QX QY QZ  TX TY TZ  CAMERA_ID  NAME
           (quaternion + translation = world-to-camera, i.e. cam_from_world)
  line 2:  X Y POINT3D_ID  X Y POINT3D_ID ...  (2D keypoints; -1 = no 3D match)
  So image count = (non-comment, non-blank lines) / 2.

points3D.txt — one line per 3D point:
  POINT3D_ID  X Y Z  R G B  ERROR  (IMAGE_ID POINT2D_IDX)...  (the track)

Binary equivalents (cameras.bin/images.bin/points3D.bin) are interchangeable;
convert with `colmap model_converter --output_type {BIN,TXT}`.

## How to run COLMAP on a scene

Set the image path per family, then:
  SCENE=south-building;  IMG=$SCENE/images            # COLMAP examples
  SCENE=relief;          IMG=$SCENE/images/dslr_images # ETH3D

Sparse (from scratch):
  colmap feature_extractor --image_path "$IMG" --database_path db.db
  colmap exhaustive_matcher --database_path db.db        # small/medium scenes
  colmap mapper --image_path "$IMG" --database_path db.db --output_path sparse
  colmap model_analyzer --path sparse/0

Reuse the prebuilt database (COLMAP example scenes only — skips extract+match):
  colmap mapper --image_path south-building/images \
    --database_path south-building/database.db --output_path sparse

Dense MVS (needs a Metal or CUDA build; after sparse):
  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

## Loading / evaluating against ground truth

- Load a model in Python: `import pycolmap; rec = pycolmap.Reconstruction("relief/dslr_calibration_jpg")`
  then `rec.cameras`, `rec.images` (each has `.cam_from_world`), `rec.points3D`.
- The ETH3D `dslr_calibration_jpg` and the COLMAP `sparse` folders ARE valid
  COLMAP models — open them directly to get GT/reference poses and points.
- Typical SfM eval: run your pipeline, align your model to the GT model
  (`colmap model_aligner` / pycolmap), then compare camera centers and rotations.

## Choosing a scene

- Fastest smoke test: relief or kicker (31 imgs each).
- Larger / stress: gerrard-hall (100) or south-building (128).
- Skip extraction+matching: the two COLMAP scenes ship database.db.
- Distortion / high-res robustness: any ETH3D scene (24 MP, fisheye).
- Pose-accuracy evaluation: any scene (all ship a GT/reference model).

## Golden MVS reference (colmap_golden_bundle/)

Also in this repo: `colmap_golden_bundle/` = the CUDA `patch_match_stereo`
golden output for the `south-building` scene (upstream COLMAP 4.0.4 + CUDA on an
NVIDIA T4). Purpose: a "did I port the SAME algorithm faithfully" reference to
diff a Metal/other MVS port against — NOT ground-truth accuracy (CUDA output is
itself an approximation; for accuracy use ETH3D/DTU real GT).

Layout:
  colmap_golden_bundle/
    note.md                              # full provenance + reader (read this)
    colmap_cuda_golden_data.ipynb        # idempotent notebook that produced it
    golden_mvs/dense/fused.ply           # 3,609,743 points (93 MB)
    golden_mvs/dense/stereo/depth_maps/*.geometric.bin   # 128 depth maps (reference)
    golden_mvs/dense/stereo/normal_maps/*.geometric.bin  # 128 normal maps
    logs/                                # undistort / patch_match / fusion logs

Map format: COLMAP dense binary — ASCII header "width&height&channels&" then
little-endian float32 in Fortran (column-major) order; 0 = invalid (ignore
zeros). Depth = HxW, normals = HxWx3. Reader:
  import numpy as np
  def read_colmap_array(path):
      with open(path, "rb") as fid:
          hdr, amp = b"", 0
          while amp < 3:
              ch = fid.read(1); hdr += ch
              if ch == b"&": amp += 1
          w, h, c = (int(x) for x in hdr.decode().split("&")[:3])
          data = np.fromfile(fid, np.float32)
      return np.transpose(data.reshape((w, h, c), order="F"), (1, 0, 2)).squeeze()

Validate a port: run `colmap patch_match_stereo` on south-building (geometric
consistency on, undistort cap 2000 px to match), then compare your
*.geometric.bin against these on overlapping valid (non-zero) pixels within
tolerance. See colmap_golden_bundle/note.md for exact pipeline + per-stage perf.

## Caveats

- The ETH3D laser-scan ground-truth surface/mesh is NOT included — only the
  GT camera calibration and GT sparse points. This is for SfM/pose and MVS
  smoke testing, not full MVS-accuracy benchmarking against the scan.
- ETH3D images are 24 MP: extraction/MVS are memory-heavy. Downscale with
  `--SiftExtraction.max_image_size 3200` (or similar) and
  `--PatchMatchStereo.max_image_size` to stay within RAM.
- THIN_PRISM_FISHEYE intrinsics must be respected; don't force a pinhole model
  on ETH3D inputs if you want the GT calibration to apply.
- These scenes are redistributed for testing; cite/respect the original
  sources (ETH3D: eth3d.net, CVPR 2017; COLMAP examples: colmap.github.io,
  CVPR 2016). See README.md.