datasetId string | heldoutPairs int64 | medianIoU float64 | visibleAnchors int64 | checkedHoldIds list | anchorInsideMaskRate float64 | anchorMisses list | holdoutUndetected int64 | note string | artifactSha256 dict |
|---|---|---|---|---|---|---|---|---|---|
9af29d404bf9e5ee | 1,488 | 0.812062 | 1,410 | [
"p007636e6b6",
"p0086765742",
"p00be02db28",
"p0107547ba5",
"p012ed519ec",
"p016d5f222b",
"p024db65c06",
"p02c77adbe1",
"p02dce5869d",
"p02f298286e",
"p034b715d53",
"p04540b4d3c",
"p047696ae20",
"p04c556b8fe",
"p0548bc2907",
"p05b331250e",
"p05ed968326",
"p0626d4923f",
"p065da2a0... | 0.959574 | [
{
"id": "p916646ccaf",
"image": "IMG_1517.jpg",
"inside": false,
"iou": 0.6094986807387863
},
{
"id": "p53f459b48f",
"image": "IMG_1517.jpg",
"inside": false,
"iou": 0.04159733777038269
},
{
"id": "p012ed519ec",
"image": "IMG_1517.jpg",
"inside": false,
"iou":... | 131 | 比較對象是獨立視角的 SAM 輪廓,非人工正解或公分量測;留出照片沒偵測到的岩點另計(holdoutUndetected)。 | {
"scene.json": "fcac906b093bf166894963f03836a87b0fd3e8ca2b223c777e4129f97c15d329",
"wall.glb": "3a36033ff2346c06eea86b5ded06029ea92ac0bc2ed76b9347053c8ea7217a58",
"face-labels.bin": "c24a5c5ae2c50642cfb4df93a051f24bbc1f495e2260cd535a24bb3e92f174cb",
"surface-mesh.npz": "8512507f2613ced061fa6eb9b88c410f9a9276ca... |
Wall3D Lab — 原岩明德 Spraywall
Photos and processing results from turning phone photos of an indoor climbing wall into a 3D wall, for one wall (the spraywall at 原岩明德). The finished wall is in the open dataset climbio/walls3d at walls3d/d1c27twdlbyor9cs/, and the camera positions here use that wall's coordinates.
原岩明德 Spraywall 一面室內岩牆的照片與處理過程,用來研究「手機拍照自動做出 3D 岩牆」。做好的牆在開放資料集 climbio/walls3d 的 walls3d/d1c27twdlbyor9cs/,這裡的相機位置用的就是那面牆的座標。
Layout / 結構
| Folder | Contents |
|---|---|
photos/originals/ |
428 original photos (iPhone 12, 2026-11-06, EXIF GPS kept) |
pipeline/ |
Processing run: 165 chosen photos resized to 900 px (images/), COLMAP sparse model and database (sparse/0/, colmap.db), Apple Object Capture mesh (apple-model/), the run's wall model (wall.glb), hold detections and per-photo segments, hold labels, alignment and validation summaries, logs |
pipeline/frames-product.json |
Camera of every photo in the coordinates of the published wall (see below) |
experiments/worldlabs-atlas/ |
World Labs Atlas test: the Gaussian-splat result (atlas.spz), photo sets tried (6, 24, 104 photos; 104 at 1280 px and 2048 px), comparison images. We kept the mesh for indoor walls |
manifest.json lists every file with size, SHA-256, license and credit. Files identical to an original photo are listed with same_as and stored once.
Camera positions / 相機位置
pipeline/frames-product.json: for each photo, R and t map a point X of the published wall (wall.json hold positions, wall-*.glb) to camera coordinates, COLMAP convention (x right, y down, z forward):
x_cam = R · X + t
pixel = f · (1 + k1·r²) · (x/z, y/z) + (cx, cy), r² = (x/z)² + (y/z)²
camera holds the COLMAP SIMPLE_RADIAL intrinsics for the 900 × 1600 images in pipeline/images/. The COLMAP model maps to the wall by X = scale · rotation · (X_colmap − center) (worldToViewer in the file). Units are arbitrary (not metres).
Checks: the 165 converted camera centres match the aligned scene to < 1e-6. Projecting the published wall's 1,122 holds into the 165 photos, 88.9 % land inside an AI hold-detection box (76.3 % after shifting all holds by 0.1 units, as a baseline).
每張照片的 R、t 把正式版牆的點換到相機座標(COLMAP 慣例)。驗證:正式版 1,122 顆岩點投影到 165 張照片,88.9% 落在 AI 偵測框內(整體平移 0.1 的對照組 76.3%)。
License and credit / 授權與署名
CC BY-SA 4.0. Credit Climbio.
Notes / 注意
- Original photos carry capture location (EXIF GPS), which is the gym. 原始照片帶拍攝位置(岩館地點)。
- Checked with a person detector: no people in the photos. 用人物偵測檢查過,照片裡沒有人。
- Local user paths in process files are replaced with
~/. 處理過程裡的本機路徑已換成~/。 - Experimental: one wall only. The processing scripts are not published yet. 實驗版:只做過一面牆,處理程式尚未公開。
- Downloads last month
- 3