---
license: cc-by-4.0
tags:
  - image-dataset
  - object-detection
  - tic-tac-toe
task_categories:
  - object-detection
---

# Tic-tac-toe Board Dataset

## Overview
- Dataset for occupancy detection of tic-tac-toe boards (top-down view).
- Contents: real images `real/images` and YOLO-format labels `real/labels`. Synthetic data can be generated on-demand with the scripts in this repo.
- Classes: 0=empty_cell, 1=white_circle_cell, 2=black_cross_cell

## License
- Dataset: CC-BY-4.0
- Real images contain no PII/portraits. Credit per README/model card.
- Source code (generation, training, inference): https://github.com/guren-kaina/AMD_Robotics_Hackathon_2025_ProjectTemplate/tree/main/mission2/code/tic_tac_toe_overlay

## Data contents
- Resolution: depends on capture device (boards are typically centered).
- Annotation: YOLO txt (class cx cy w h) normalized coords. Matching `.txt` lives in `real/labels`.
- Synthetic data: generated via `main.py --num-train/--num-val` (noise/contrast changes/distractors included) into `data/synth_grid/`.
- Parquet export: `dataset.parquet` (generated by `make hf-dataset-stage`) with columns: image, label_file, class_id, cx, cy, w, h. One row per bbox; empty label files produce a single row with null class/coords.

## Suggested splits
- Real images are limited; mix synthetic:real around 8:2 and hold out val/test separately.

## Notes and care
- Optimized for top-down view; for oblique viewpoints or different board designs, re-label and retrain.
- For new token designs or colors, add annotations and retrain.

## Usage example
```bash
python main.py --real-data real --num-train 1000 --num-val 200 --force-train
```