nithinmanoj10/CatanSynth-Terrain-v1-50K

CatanSynth Terrain v1 (50K)

Procedurally-generated synthetic patches of individual Settlers of Catan hexagonal terrain tiles, each pre-rectified to a top-down 224×224 crop and labeled with terrain type and resource number token. Generated by the same `catan-in-silico` synthetic data engine as CatanSynth Meridian v1 (50k), but composed of pre-cropped single-tile patches rather than full board images. Used to train CataNET-Terrain-v1.0-Small and CataNET-Terrain-v1.0-Large, the terrain/token classification stage of the CataNET pipeline.

Despite the "50K" name, the actual image count is 51,984 — a legacy of the dataset's naming history; see Splits for exact counts.

Task

Two-label classification per image, one hex tile per patch (not a multi-instance/detection task, unlike CatanSynth Meridian):

• —Terrain (6 classes): desert, fields, forest, hills, mountains, pasture
• —Token (12 classes): none (desert only), robber (robber-occupied, any terrain), or the tile's number 2–6, 8–12 (7 is never a valid tile number in Catan and is excluded from this dataset's labels — the downstream model architecture reserves an unused output slot for it, but no training example ever has it)

Each image is one already-isolated, top-down-warped tile — no bounding boxes or keypoints, since the geometry problem is solved by the separate CatanSynth Meridian dataset/model pair. This is a dual-head classification dataset, not a detection one.

Splits

Each split's images are chunked into chunk_000/, chunk_001/, … subfolders of ≤8,000 files each, to stay under Hugging Face's recommended ~10k-files-per-folder limit (train spans 6 chunks; val/test fit in one each).

Format

Every filename fully encodes its own ground truth: <terrain>_<token>_<board_id>_<tile_idx>.jpg, e.g. mountains_11_00099_4.jpg = terrain mountains, token 11, from board 00099, tile index 4. Two derived, parallel representations are also included for convenience:

1. 1.`images/labels.csv` — one row per image (filename, terrain, token, split), for use with a plain pandas/CSV-based dataloader (this is a drop-in replacement for the schema the original training script expects — point its DATA_DIR at the downloaded images/ folder and it works unmodified).
2. 2.`metadata.jsonl` (one per split, next to images/<split>/) — for the Hugging Face imagefolder loader / Dataset Viewer, adding terrain, token, terrain_id, token_id, board_id, tile_idx as structured, typed columns (the terrain/token columns are declared as class_label in this card's metadata, which is what gets the Hub to render categorical distribution charts in the Viewer rather than plain filterable text).

Loading

from datasets import load_dataset

ds = load_dataset("nithinmanoj10/CatanSynth-Terrain-v1-50K")
ds["train"][0]  # {"image": <PIL.Image>, "terrain": "mountains", "token": "11", ...}

To train directly with the original dual-head MobileNetV3 training script, download the repo and point DATA_DIR at the images/ folder — images/labels.csv matches the schema it expects.

Relationship to CatanSynth Meridian

This dataset and CatanSynth Meridian v1 (50k) come from the same underlying synthetic board renders but serve different stages of the CataNET pipeline: Meridian's images are full boards labeled for hex-corner keypoint detection (pose estimation), while this dataset's images are individual, already-rectified tiles labeled for terrain/token classification. They are not interchangeable and are not guaranteed to be sampled from exactly the same underlying board generations.

License

MIT