ChessRender360: High-Fidelity Rendered Chess Dataset with Multi-Modal Annotations

ChessRender360 is a synthetically crafted dataset featuring 10,000 rendered chess positions. Designed for computer vision and machine learning research, this dataset provides a rich collection of RGB images, depth maps, instance masks, and semantic segmentation masks for each chess piece and board element.

Each chess position is rendered in high resolution (2000x2000 pixels), capturing the intricate details of the board and pieces from various angles. The dataset includes:

RGB Images: High-quality rendered images of chess positions, showcasing a diverse range of board configurations.

Depth Maps: Accurate depth representations of the scene, capturing depth in the range of 20 cm to 120 cm. In the depth maps, black corresponds to a depth of 20 cm, and white corresponds to 120 cm, providing spatial information for each position.

Instance Masks: Unique instance masks for each chess piece, enabling precise identification and localization.

Semantic Segmentation Masks: Segmentation masks that differentiate between piece types and board elements, with distinct hue values assigned to each type.

Bounding Boxes: Each sample has an annotation .json file containing bounding boxes for each piece.

Board Corners: Same annotation .json contains positions of corners of the board in order: white left, white right, black left, black right.

FENs: A CSV file containing the FEN (Forsyth-Edwards Notation) for each chess position in the dataset, listed in order. This allows users to easily recognize and replicate the exact board position from any image.

Bounding Box Generation:

The dataset does not include predefined bounding boxes, but they can be easily generated from the provided semantic and instance masks. This allows for flexible bounding box creation tailored to specific research needs.

Rendering Details:

3D Models and Materials: The dataset uses a consistent set of 3D models for all chess pieces, with three different material/color schemes applied across the dataset, along with random perturbations in material brightness, contrast and saturation to introduce visual variety.

Camera Angles: Camera angles are randomly selected, with yaw ranging from 0 to 360 degrees and pitch between 30 to 80 degrees, providing diverse perspectives of the chess positions.

Background Variability: The chessboard is randomly placed on different types of tables, with the floor material randomly sampled to create a variety of backgrounds.

Lighting: Lighting conditions are randomly generated, adding further diversity and realism to the rendered scenes.

Augmentation Potential:

The instance and semantic masks can be used to further augment the dataset. Researchers can selectively modify specific parts of the images—such as the board, background, or individual pieces—enabling the creation of new variations and enhancing the dataset's utility for model training and testing.

Color Mapping:

The semantic masks are color-coded using a hue-based system, where the board frame, squares, and each piece type are assigned specific hues. Instances of the same piece type are differentiated by varying the value component, with saturation consistently set to 1. A detailed color_mapping.json file is included, providing a comprehensive guide to interpreting the masks.

Applications: ChessRender360 is ideal for tasks such as object detection, instance segmentation, depth estimation, and scene understanding in synthetic environments. Researchers and developers can leverage this dataset for training and evaluating models in computer vision, robotics, and artificial intelligence.

Dataset Highlights:

10,000 uniquely rendered chess positions

High-resolution images (2000x2000 pixels) with diverse visual characteristics

Comprehensive annotations with RGB, depth (20 cm to 120 cm), instance, and semantic maps

Side identification map for distinguishing white and black sides of the board

FEN notation CSV file for easy position recognition and replication

Variety introduced through different material schemes and lighting setups

Randomized camera angles for enhanced perspective diversity

Potential for further augmentation by modifying specific image components

Detailed color mapping for easy interpretation of segmentation masks

Suitable for a wide range of computer vision and AI applications

ChessRender360 offers a rich and versatile dataset for advancing research and development in the field of computer vision, providing a synthetic yet highly realistic environment for model training and testing.

For any questions, feedback, or collaboration opportunities, or if you are interested in custom artificial datasets, please contact me:

Name: Marko Kojić

LinkedIn: https://www.linkedin.com/in/mmkoya

I welcome inquiries from researchers, developers, and organizations interested in utilizing or collaborating on artificial datasets.