Dataset Card for Blackjack

Dataset Description

Dataset Summary

A dataset containing two sets of playing card images for hands in the card game Blackjack. Each set contains at least 10,000 images and has a series of attributes. This dataset is based on the dataset Playing cards [1]

Train and test splits are provided in both JSON and pickle formats. Concept and task classification labels (both zero indexed) and names are provided in txt files.

Dataset Structure

Data Instances

Each set of samples have the following:

• player and dealer playing cards in each sample image
• A list of concepts present in the each sample (1 for concepts present and 0 otherwise)
• The task classification label
• coordinates for each of the corners of playing cards in each sample.

The basic structure of the JSON and pkl files describing each sample is as follows:

sample ID, {
    'img_path': string file path,
    'class_label': integer,
    'concept_label': list of 0s and 1s,
    'player_card_points': list of tuples and card class labels as integers
    'dealer_card_points': list of tuples and card class labels as integers
    'game_numer': integer
}

Standard

Card hands using a single style of playing cards.

• Concepts: soft/hard hand, sum of player cards, first dealer card, dealer has multiple cards
• Class label: Best move
• Card points: Coordinates of the card and card classification

Example

"14304": {
  "img_path": "imgs/standard/val/0/14304.png",
  "class_label": 0,
  "concept_label": [0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0],
  "player_card_points": [[[[50, 789], [173, 789], [50, 974], [173, 974]], "QS"], [[[185, 789], [308, 789], [185, 974], [308, 974]], "5S"]],
  "dealer_card_points": [[[[172, 235], [50, 235], [172, 50], [50, 50]], "7D"]],
  "game_number": 0
}

Mixed

Card hands using a one style of playing cards for all Ace and Seven playing cards and a second style for all other cards.

• Concepts: soft/hard hand, sum of player cards, first dealer card, dealer has multiple cards
• Class label: Best move
• Card points: Coordinates of the card and card classification

Example

"0": {
  "img_path": "imgs/mixed_ace_seven/train/0/0.png",
  "class_label": 0,
  "concept_label": [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0],
  "player_card_points": [[[[173, 974], [50, 974], [173, 789], [50, 789]], "10S"], [[[185, 789], [308, 789], [185, 974], [308, 974]], "4H"]],
  "dealer_card_points": [[[[172, 235], [50, 235], [172, 50], [50, 50]], "QC"]],
  "game_number": 0
}

Data Fields

• String file path from the root of the dataset to a given samples image file
• A list of concepts present in the each sample (1 for concepts present and 0 otherwise). The index of each value in this list corresponds to the label in concepts.txt.
• The task classification label. This corresponds the the label in classes.txt
• list of playing cards present in a given sample player hand. Each item in the list has a list of card coordinates (card coordinates are always in the order top left, top right, bottom left, bottom right) and the card classification label.
• list of playing cards present in a given sample player hand. Each item in the list has a list of card coordinates (card coordinates are always in the order top left, top right, bottom left, bottom right) and the card classification label.
• A number representing the game the sample belongs to. Samples are in order with full games of backjack represented.

Data Splits

Standard

Task classifications

Class name	Count train	Count val
hit	3576	1554
stand	3576	1554
surrender	3576	1554
bust	3576	1554

Concepts

Concept name	Count train	Count val
soft	869	325
hard	13435	5891
player_value_21_plus	3576	1554
player_value_21	620	278
player_value_20	714	326
player_value_19	517	220
player_value_18	554	235
player_value_17	621	270
player_value_16	3994	1720
player_value_15	724	271
player_value_14	624	245
player_value_13	599	269
player_value_12	591	270
player_value_11	306	165
player_value_10	215	108
player_value_9	192	85
player_value_8	457	200
dealer_card_2	735	373
dealer_card_3	750	347
dealer_card_4	810	317
dealer_card_5	791	339
dealer_card_6	821	351
dealer_card_7	989	343
dealer_card_8	901	321
dealer_card_9	859	411
dealer_card_10	6119	2773
dealer_card_a	1529	641
dealer_multi_cards	1788	778

Mixed

Task classification

Class name	Count train	Count val
hit	3558	1550
stand	3558	1550
surrender	3558	1550
bust	3558	1550

Concepts

Concept name	Count train	Count val
soft	849	343
hard	13383	5857
player_value_21_plus	3558	1550
player_value_21	621	260
player_value_20	705	308
player_value_19	568	255
player_value_18	542	236
player_value_17	555	240
player_value_16	3982	1741
player_value_15	709	286
player_value_14	655	276
player_value_13	617	259
player_value_12	556	277
player_value_11	292	112
player_value_10	219	107
player_value_9	206	92
player_value_8	447	201
dealer_card_2	832	349
dealer_card_3	787	327
dealer_card_4	813	372
dealer_card_5	720	358
dealer_card_6	774	324
dealer_card_7	841	367
dealer_card_8	804	388
dealer_card_9	875	375
dealer_card_10	6370	2711
dealer_card_a	1416	629
dealer_multi_cards	1783	776

Dataset Creation

Curation Rationale

This dataset was created to test Concept Bottleneck Models [2] in a human-machine setting.

Source Data

Initial Data Collection and Normalization

The dataset uses background from [3] and playing card images from [4]. The dataset is balanced to the task classification labels. The code used to generate the dataset is available here [5].

Annotations

Annotation process

The annotation process was completed during the generation of the dataset.

Who are the annotators?

Annotations were completed by a machine.

Personal and Sensitive Information

This dataset does not contain personal and sensitive Information.

Additional Information

Licensing Information

This dataset is licenced with the MIT licence.

Citation Information

[1] Furby, J., Cunnington, D., Braines, D., Preece, A.: Can we constrain concept bottleneck models to learn semantically meaningful input features? (2024), https://arxiv.org/abs/2402.00912

[2] Koh, P.W., Nguyen, T., Tang, Y.S., Mussmann, S., Pierson, E., Kim, B. & Liang, P.. (2020). Concept Bottleneck Models. Proceedings of the 37th International Conference on Machine Learning, in Proceedings of Machine Learning Research 119:5338-5348 Available from https://proceedings.mlr.press/v119/koh20a.html.

[3] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed and A. Vedaldi, "Describing Textures in the Wild," 2014 IEEE Conference on Computer Vision and Pattern Recognition, 2014, pp. 3606-3613, doi: 10.1109/CVPR.2014.461.

[4] j4p4n, "Full Deck Of Ornate Playing Cards - English", Available at: https://openclipart.org/download/315253/1550166858.svg

[5] J. Furby, "blackjack-dataset-generator", Available at: https://github.com/JackFurby/blackjack-dataset-generator