GeoSetEnnea: A Multi-Class Image Dataset for Object Detection of Geometric Tools

GeoSetEnnea is a curated image dataset developed to support object detection research focused on geometric tools commonly used in educational, design, and industrial settings. The dataset was created to address the gap in publicly available resources representing such tools, which are often used in academic and technical settings but are underrepresented in computer vision studies.

The dataset contains 966 annotated images with 8253 annotations of 9 different classes of geometric tools, including compasses, protractors, erasers, set squares etc. Images were captured using three different smartphone models under various lighting conditions, angles, and backgrounds across several stationery shops and libraries in Dhaka, Bangladesh. This diversity enhances the generalizability and robustness of object detection models trained on this dataset.

All images were annotated with bounding boxes using the Roboflow platform and are split into training (70%), validation (20%), and testing (10%) sets. The dataset includes a structured folder hierarchy and metadata files to facilitate easy integration into object detection workflows. Preprocessing steps such as auto-orientation, resizing (640×640), and data augmentation (e.g., flips, brightness, contrast, gamma correction) were applied to improve dataset quality.

Our hypothesis is that a diverse, real-world dataset of shape-similar tools can significantly improve model performance in practical detection scenarios. This is validated by training 4 state-of-the-art models: YOLOv7, YOLOv8, YOLOv11, and RF-DETR on GeoSetEnnea. RF-DETR achieved the highest performance with a mAP@50 of 99.6%, demonstrating the dataset's reliability and effectiveness.

Researchers and developers can use GeoSetEnnea to train and benchmark detection systems for educational applications (e.g., virtual classrooms), mobile learning tools, or inventory management systems. The dataset structure and documentation ensure clarity on data origin, labeling, and intended usage, supporting reproducible and impactful research.