WildFuria/solitaire-cards-dataset

Solitaire Cards Dataset

Images of playing cards captured from live Microsoft Solitaire Collection games (1920×1040 window, UWP app) for training the vision layer of an autonomous computer-use agent.

Configurations

Folder	Images	Size	What it is
full	3254	131×176 px	complete face-up cards — the training set of the published recognizer
corners	3967	42×54 px	corner crops from the earlier, abandoned pipeline (kept for comparison)
fixtures	3	1920×1040 px	real screenshots used as regression tests, with a hand-read ground truth

All 52 classes are stored as folders named <rank>_<suit> (A_hearts, 10_spades, 2_clubs, …, K_diamonds), which is exactly the class order torchvision.datasets.ImageFolder produces (alphabetical) and the order of the published model's outputs.

How it was collected

1. scripts/collect_full.py runs while a game is open and saves every fully visible card crop to dataset_full/_unlabeled/.
2. scripts/sort_full.py asks a local vision-language model (qwen2.5vl:7b via Ollama) for the rank and suit, rejects images that are actually a stack of two cards, and moves the rest into <rank>_<suit>/ folders.
3. Ambiguous cases were sorted manually.

Label noise is expected. The labels come from a model plus manual sorting, so the 99.91 % accuracy of the published recognizer is measured against these labels, not against a human-audited gold standard.

Ground truth for the fixtures

fixtures/ contains three real screenshots with a manually read table state (written down card by card, including the rank and suit of every visible card):

• live_frame_2026-10-01.png — midgame with deep fans (columns 2 and 6 empty)
• live_frame_empty_cols_2026-10-01.png — the same game later, when columns 2 and 6 are empty
• live_frame_fresh_deal_2026-10-01.png — a fresh deal, used to catch the "blue portrait mistaken for a card back" bug (in it the bottom card of column 5 is Q♣, whose dress is drawn in blue)

The exact expected layout per frame is in tests/test_screen_to_board_real_frame.py.

Intended use

Training and benchmarking card classifiers and, more generally, perception layers for computer-use agents. Not suitable for anything that requires human-audited labels.

Links

• Agent code: solitaire-computer-agent
• Recognizer: solitaire-card-recognizer
• Case study: ai-wiki.tech

License and provenance

MIT for the dataset packaging. The images show the Microsoft Solitaire Collection card art and are published for research and educational purposes; this dataset is not affiliated with or endorsed by Microsoft.