CubeCheck renders

180,200 synthetic 3D renders of scrambled 2×2 to 7×7 cubes, and of rooms with no cube, each with exact labels: the cube's seven visible corners, every sticker's colour and position, and the scramble. They trained the CubeCheck models, which check a scrambled cube from one camera view at check.cubing.in.

• Models: darshanvss/cubecheck (c7 corners and the r4 sticker reader were trained on this data)
• Code, training and evaluation: github.com/darshanvss/cubecheck

Subsets and splits

One subset per puzzle, each with a train and a test split. Test sets never share a scene with train sets: a random seed fixes a whole scene, and no two sets share seeds, except none/train-v1, which is the first 5,000 rooms of 333/train-full-v1 with the cube hidden (both train).

Subset	Train	Test	Image sizes
2x2	20,000	1,900	720×720, 960×720, 1024×768, 1280×720
3x3	51,000	3,400	720×720, 960×720, 1024×768, 1280×720
4x4	20,000	1,900	720×720, 960×720, 1024×768, 1280×720
5x5	20,000	1,900	720×720, 960×720, 1024×768, 1280×720
6x6	20,000	1,900	720×720, 960×720, 1024×768, 1280×720
7x7	20,000	1,900	720×720, 960×720, 1024×768, 1280×720
no-cube	15,000	1,300	720×720, 960×720, 1024×768, 1280×720

from datasets import load_dataset

ds = load_dataset("darshanvss/cubecheck", "4x4", split="test")
ds[0]["jpg"]          # a PIL image
ds[0]["json"]["set"]  # "444/test-mixed-v2"

# One kind of view only, e.g. the 4x4's two-face test views:
two_faces = load_dataset("darshanvss/cubecheck", data_files="data/444/test-shallow-v2-*.tar", split="train")

Each subset mixes several kinds of view; the set field in every label (and the file name of its shard) says which:

Kind	What's in it
mixed	three faces toward the camera, scrambles of every length (solved to a full scramble)
full, easy (3×3)	full 25-move scrambles; solved or nearly solved cubes
shallow	two faces and a sliver of the third, the cube held toward the scrambler
edgeon	exactly two faces, an edge toward the camera
small	the cube far away, few pixels a sticker
busy (3×3)	cluttered scenes
no-cube	rooms, tables, hands and other objects with no cube: for learning to say "no cube here"

v2/v3 are later rounds of the same kind with new seeds. Each set's settings (seeds, JPEG quality, how many are handheld) are in manifests/<puzzle>/<set>.json.

Files

data/<puzzle>/<set>-NNNN.tar    WebDataset shards of about 1 GB: syn_0000001.jpg next to syn_0000001.json
manifests/<puzzle>/<set>.json   each set's settings
sources.json                    the 117 Poly Haven assets the scenes use, with their authors

<puzzle> is the WCA event id (222 … 777), none for no cube. To get the files back as folders of images and labels, as CubeCheck's training code reads them:

mkdir -p 444/test-mixed-v2/images 444/test-mixed-v2/labels
tar -xf data/444/test-mixed-v2-0000.tar -C 444/test-mixed-v2/images --wildcards '*.jpg'
tar -xf data/444/test-mixed-v2-0000.tar -C 444/test-mixed-v2/labels --wildcards '*.json'

Labels

One JSON per render (abridged; a 3×3 has 54 stickers, an N×N 6N²):

{
  "id": "syn_68000284", "image": "images/syn_68000284.jpg", "set": "777/train-edgeon-v3",
  "seed": 68000284, "mode": "handheld", "width": 1024, "height": 768,
  "scramble": "3Lw Rw' 3Dw' 3Lw' 3Uw U' Bw2 ...",
  "cube": {"size": 7, "finish": "glossy", "type": "sticker", "palette": "vibrant", "logo": false, "orientation": [...]},
  "camera": {"fov": 45.65, "distance": 29.07, "roll": -1.78, "position": [...], "quaternion": [...]},
  "lighting": {"keyKelvin": 3241, "environment": "panorama", "exposure": 1.09, "shadows": false, ...},
  "background": {"kind": "panorama", "source": "pav_studio_02.jpg", "table": null},
  "occluders": {"fingers": 2, "frontFingers": 1, "palm": false},
  "labels": {
    "facelets": "DURRBDURBURBFF...",
    "keypoints": [[563.81, 401.23, 2], ...],
    "visibleFaces": ["D", "L", "F"],
    "bbox": [390.5, 379.77, 702.1, 655.0],
    "corners": [{"name": "UFR", "x": 394.11, "y": 611.13, "inFrame": true, "occluded": true}, ...],
    "stickers": [{"facelet": 0, "color": "D", "x": 681.06, "y": 611.63, "facing": -0.14, "inFrame": true, "visibleFraction": 0, "visible": false}, ...]
  }
}

• facelets: the whole cube's sticker colours, face by face in U R F D L B order, row by row (WCA colours: U white, R red, F green, D yellow, L orange, B blue). It's the state the scramble gives, in the cube's own orientation.
• keypoints: the seven visible corners as [x, y, v] in pixels, the six around the outside clockwise and then the one in the middle (v: 2 in view, 1 covered, 0 outside the frame, with x and y 0). In an exactly-two-face view both ends of the shared edge are fair answers for the middle one.
• corners: all eight corners by name (UFR, …), in pixels, with whether each is in the frame and whether it's covered.
• stickers: every sticker's centre in pixels, whether it faces the camera, and how much of it is visible (fingers and the cube's own edges hide some).
• mode: handheld (held in rendered fingers) or table.
• No-cube renders have "empty": true and empty corners and stickers.

How it was made

Scenes were built and rendered with a 3D renderer of our own (not published): a cube of a random size, finish, sticker style and colour shades; a random scramble; a camera of random lens, distance and roll; lights of random colour temperature; and a background from 60 panoramas and 57 table textures, all CC0 photos from Poly Haven (sources.json). Some cubes are held in rendered fingers. The images contain no photographs of people.

The renders are clean: camera effects (blur, noise, compression, colour casts) and painted fingers over the corners and edges are added at training time by CubeCheck's training code, not baked into the images.

Limits

• Synthetic only. Models trained on it need checking on real camera frames.
• Standard WCA colour scheme, sticker and stickerless-style cubes; no other colour schemes, no other puzzles yet (pyraminx, skewb and the rest are planned as new subsets).

Licence

The renders and labels are CC BY 4.0: use them for anything, crediting CubeCheck (link to this page or to github.com/darshanvss/cubecheck). The background photos are CC0 from Poly Haven; their authors are listed in sources.json. The CubeCheck code and models are AGPL-3.0.

Versions

Each release is a tag. v1.0: the 2×2 to 7×7 renders and the no-cube scenes that trained c7 and r4.