Datasets:

vietanhdev
/

chessai-data

pretty_name: ChessAI Data — Chinese Chess Piece Detection
language:
  - vi
size_categories:
  - 1K<n<10K
task_categories:
  - object-detection
tags:
  - chess
  - chinese-chess
  - xiangqi
  - co-tuong
  - object-detection
  - bounding-box
  - anylabeling
authors:
  - Viet-Anh Nguyen

ChessAI Data — Chinese Chess Piece Detection

A bounding-box detection dataset for Chinese Chess (Cờ tướng / 象棋), labeled with AnyLabeling. Built to train piece-recognition models for chessboard-state extraction from photos.

Dataset summary

Total annotated images: 1,747 (per-image AnyLabeling JSON in data/combined_data/)
COCO-format split: 872 images, 18,790 bounding boxes (data/annotations.json)
Classes: 7 (the standard Xiangqi piece set)
Total size: ~280 MB (images + annotations)

Classes

Vietnamese piece names are used throughout. Counts below are from the COCO split.

ID	Label (VN)	Piece	Boxes
1	`xe`	Chariot (rook)	2,264
2	`ma`	Horse (knight)	2,357
3	`tuong`	Elephant (bishop-like)	2,350
4	`si`	Advisor (palace guard)	2,375
5	`vua`	General (king)	1,244
6	`phao`	Cannon	2,369
7	`tot`	Soldier (pawn)	5,831

Files

data/combined_data/ — paired .jpg + .json files in labelme / AnyLabeling format. Each .json has a shapes[] array with label, points (top-left and bottom-right corners), and shape_type: "rectangle".
data/annotations.json — COCO-format export covering 872 images and 18,790 boxes, ready for use with detection libraries that expect COCO.
data/data_01/, data/data_02/ — raw images grouped by capture session.
make_data.sh — pipeline that produces the COCO export from the raw + per-image annotation pairs.

Quick start — load the COCO split

from huggingface_hub import hf_hub_download
import json

path = hf_hub_download(
    repo_id="vietanhdev/chessai-data",
    filename="data/annotations.json",
    repo_type="dataset",
)
with open(path) as f:
    coco = json.load(f)

print(len(coco["images"]), "images,", len(coco["annotations"]), "boxes")
print("classes:", [c["name"] for c in coco["categories"]])

To download images alongside, use snapshot_download with allow_patterns=["data/combined_data/*"].

Reproducing the dataset

conda create -n chessai-dataprep python=3.9
conda activate chessai-dataprep
pip install -r requirements.txt
# After labeling raw images in data/data_01 and data/data_02 with AnyLabeling:
bash make_data.sh

Source code

Upstream repo with preprocessing scripts: https://github.com/nrl-ai/chessai-data

Citation

@misc{nguyen2024chessai,
  author    = {{Viet-Anh NGUYEN (Andrew)}},
  title     = {ChessAI Data — Chinese Chess piece detection dataset},
  year      = {2024},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/2812},
  url       = {https://huggingface.co/datasets/vietanhdev/chessai-data}
}