Datasets:
Document colmap_golden_bundle in llm.txt
Browse files
llm.txt
CHANGED
|
@@ -96,6 +96,42 @@ Dense MVS (needs a Metal or CUDA build; after sparse):
|
|
| 96 |
- Distortion / high-res robustness: any ETH3D scene (24 MP, fisheye).
|
| 97 |
- Pose-accuracy evaluation: any scene (all ship a GT/reference model).
|
| 98 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 99 |
## Caveats
|
| 100 |
|
| 101 |
- The ETH3D laser-scan ground-truth surface/mesh is NOT included — only the
|
|
|
|
| 96 |
- Distortion / high-res robustness: any ETH3D scene (24 MP, fisheye).
|
| 97 |
- Pose-accuracy evaluation: any scene (all ship a GT/reference model).
|
| 98 |
|
| 99 |
+
## Golden MVS reference (colmap_golden_bundle/)
|
| 100 |
+
|
| 101 |
+
Also in this repo: `colmap_golden_bundle/` = the CUDA `patch_match_stereo`
|
| 102 |
+
golden output for the `south-building` scene (upstream COLMAP 4.0.4 + CUDA on an
|
| 103 |
+
NVIDIA T4). Purpose: a "did I port the SAME algorithm faithfully" reference to
|
| 104 |
+
diff a Metal/other MVS port against — NOT ground-truth accuracy (CUDA output is
|
| 105 |
+
itself an approximation; for accuracy use ETH3D/DTU real GT).
|
| 106 |
+
|
| 107 |
+
Layout:
|
| 108 |
+
colmap_golden_bundle/
|
| 109 |
+
note.md # full provenance + reader (read this)
|
| 110 |
+
colmap_cuda_golden_data.ipynb # idempotent notebook that produced it
|
| 111 |
+
golden_mvs/dense/fused.ply # 3,609,743 points (93 MB)
|
| 112 |
+
golden_mvs/dense/stereo/depth_maps/*.geometric.bin # 128 depth maps (reference)
|
| 113 |
+
golden_mvs/dense/stereo/normal_maps/*.geometric.bin # 128 normal maps
|
| 114 |
+
logs/ # undistort / patch_match / fusion logs
|
| 115 |
+
|
| 116 |
+
Map format: COLMAP dense binary — ASCII header "width&height&channels&" then
|
| 117 |
+
little-endian float32 in Fortran (column-major) order; 0 = invalid (ignore
|
| 118 |
+
zeros). Depth = HxW, normals = HxWx3. Reader:
|
| 119 |
+
import numpy as np
|
| 120 |
+
def read_colmap_array(path):
|
| 121 |
+
with open(path, "rb") as fid:
|
| 122 |
+
hdr, amp = b"", 0
|
| 123 |
+
while amp < 3:
|
| 124 |
+
ch = fid.read(1); hdr += ch
|
| 125 |
+
if ch == b"&": amp += 1
|
| 126 |
+
w, h, c = (int(x) for x in hdr.decode().split("&")[:3])
|
| 127 |
+
data = np.fromfile(fid, np.float32)
|
| 128 |
+
return np.transpose(data.reshape((w, h, c), order="F"), (1, 0, 2)).squeeze()
|
| 129 |
+
|
| 130 |
+
Validate a port: run `colmap patch_match_stereo` on south-building (geometric
|
| 131 |
+
consistency on, undistort cap 2000 px to match), then compare your
|
| 132 |
+
*.geometric.bin against these on overlapping valid (non-zero) pixels within
|
| 133 |
+
tolerance. See colmap_golden_bundle/note.md for exact pipeline + per-stage perf.
|
| 134 |
+
|
| 135 |
## Caveats
|
| 136 |
|
| 137 |
- The ETH3D laser-scan ground-truth surface/mesh is NOT included — only the
|