Digital Chessboard Workflow - Training Data

Image Size, Cropping & Format

For optimal training quality, all images in our dataset will be cropped to 800×800 pixels before annotation and saved in JPEG (.jpg) format at 90–95% quality. This ensures that YOLO receives the same type of input during both training and real-world inference. We have developed a Chessboard Cropping Tool, available in our GitHub repository, to automate this process: 🔗 ChessboardDetector GitHub Repository

Why Crop the Images Before Training?

• Consistency with App Workflow: Our app pre-crops the chessboard before running YOLO. Training on full-screen images would introduce unnecessary variations (UI elements, glare, side menus), making detection harder.
• Better Model Accuracy: By removing irrelevant parts of the image, YOLO can focus entirely on piece detection rather than board localization.
• More Efficient Training: Smaller, focused images lead to faster training times and lower memory usage, improving overall performance.

Best Image Format for Roboflow & YOLO

For optimal training performance, file size efficiency, and preserving image quality, we use:

Format	Pros	Cons	Best Use Case
JPEG (.jpg) ✅	✔ Small file size ✔ Fast training speed ✔ Good quality if saved at 90%+	❌ Slight compression loss	Best for YOLO training & Roboflow uploads
PNG (.png) ⚠️	✔ Lossless quality ✔ Supports transparency	❌ Larger file sizes ❌ Slower training	Not necessary for YOLO
WebP (.webp) 🟡	✔ High compression + quality ✔ Smaller than JPEG	❌ Not natively supported by YOLO	Not recommended
TIFF (.tiff) ❌	✔ Lossless quality ✔ High bit depth	❌ Huge file sizes ❌ Slows down training	Overkill for YOLO

Example Image Properties

Below is an example of a properly formatted image for training YOLO and uploading to Roboflow. This image meets all dataset requirements:

✅ Format: JPEG (.jpg)
✅ Image Size: 800×800 pixels
✅ Color Model: RGB
✅ File Size: ~100-250 KB (efficient for processing)
✅ DPI: 72 pixels/inch (not an issue for YOLO training)

---

Recommended Dataset Size

To achieve our target accuracy (≥ 95% chessboard detection and ≥ 90% piece detection), our dataset should be split as follows:

Dataset Split	Minimum Recommended	Ideal
Training Set	700	800
Validation Set	200	150
Test Set	100	50
Total	1000 images	1000 images

Dataset Composition

Research & Justification for Dataset Proportions

To ensure our dataset represents real-world usage, we conducted a deep analysis of Chess.com’s most popular board themes. Based on Chess.com user data, forum discussions, and community insights, we found that over 90% of players use a small selection of predefined board themes. Our dataset proportions are directly based on this research:

• Classic Green (35%): The most widely used board, the default choice for most users.
• Brown (20%): A widely chosen flat brown theme mimicking wood, using "Classic" pieces.
• Dark Wood (15%): A highly popular wood-textured theme.
• Blue (8%): A commonly used alternative for contrast and visibility.
• Icy Sea (8%): A sleek, modern choice, preferred by blitz players.
• Tournament Dark Green (5%): Resembles classic tournament boards.
• High-Contrast B&W (3%): Ensures visibility under extreme contrast conditions.
• Glass (3%): A niche but recognized transparent board design.
• Bubblegum (2%): A vibrant, less serious choice used by streamers and casual players.
• Lolz (1%): A rare, high-contrast black/grey board that presents unique challenges.

Final Dataset Breakdown

Theme Name	# Images	Source Type
Classic Green (default)	350	Camera (280), Screenshots (70)
Brown (Brown board, Classic Pieces)	200	Camera (160), Screenshots (40)
Icy Sea	80	Camera (65), Screenshots (15)
Dark Wood (Walnut)	150	Camera (120), Screenshots (30)
High-Contrast (B&W)	100	Camera (80), Screenshots (20)
Tournament (Dark Green)	50	Camera (40), Screenshots (10)
Glass Pieces	30	Camera (24), Screenshots (6)
Bubblegum	20	Camera (16), Screenshots (4)
Lolz (Dark/grey)	10	Camera (8), Screenshots (2)
Total	1000

Labeling Strategy

• We will label individual chess pieces with bounding boxes on cropped chessboard images.
• We are using Roboflow for annotation, which allows us to efficiently label and manage our dataset while ensuring consistency across different board themes and piece types.
• Our labels will include:
  • white_pawn, white_rook, white_knight, white_bishop, white_queen, white_king
  • black_pawn, black_rook, black_knight, black_bishop, black_queen, black_king

Bias Prevention Strategy

• 1️⃣ Use a Mix of Opening, Midgame, and Endgame Positions: Ensure dataset includes game states where some pawns are missing (not just opening positions).

• 2️⃣ Include Tactics & Studies: Some datasets can include puzzle positions, sacrifices, and pawnless endgames to balance things.

• 3️⃣ Filter Out Too Many “Full Pawn Ranks” Boards: If too many images have 8 pawns per side, balance it by adding captured-pawn scenarios.

Data Augmentation

We will use augmentation techniques to automatically simulate image quality variations such as:

• Perspective transformations (angled photographs)
• Brightness and contrast adjustments (lighting variations)
• Blur and noise (lower-quality captures)

This effectively increases our dataset size by approximately 3–4 times, significantly enhancing model generalization.

Augmentation Pipeline (Python example)

import albumentations as A
import cv2

augmentation_pipeline = A.Compose([
    A.Perspective(scale=(0.05, 0.1), p=0.6),
    A.RandomBrightnessContrast(p=0.7),
    A.GaussianBlur(blur_limit=(3, 7), p=0.4),
    A.GaussNoise(var_limit=(10, 30), p=0.3),
    A.Rotate(limit=10, p=0.5),
    A.RandomCrop(width=700, height=700, p=0.2),
])

# Usage example:
image = cv2.imread('original_image.png')
augmented = augmentation_pipeline(image=image)['image']
cv2.imwrite('augmented_image.png', augmented)

Roles and Responsibilities

Our Role	Augmentation’s Role
Capture and label original Chess.com images	Generate realistic image variations
Set augmentation parameters	Automatically simulate quality variations
Periodically validate augmented images	Efficiently create diverse training examples

Next Steps

• Collect our initial images based on the above table (total 1000).
• Set up the augmentation pipeline using Albumentations.
• Train the YOLO model and export it to CoreML for integration into our iOS application.
• Continuously validate the model’s performance and improve based on user feedback.