Descriptor: Geometric Breaks Dataset (GB)

With the rise in research investigating repair and reassembly of fractured objects using deep learning, large datasets that enable training algorithms for assembly and repair from 3D models have become necessary. This paper describes the Geometric Breaks dataset, a large repository consisting of 3D models of 25,249 broken object models and their repair counterpart fragments. The broken models are generated by subtracting fracturing shapes from 22,165 complete 3D object models, acquired from 8 classes in the ShapeNet dataset and from the Google Scanned Objects dataset. Since the subtraction process generates multiple fragments, we share individual 3D models of the fragmented parts. The Geometric Breaks dataset has value in enabling training of algorithms for repair and reassembly, as well as for traditional tasks in vision and robotics related to object analysis, such as category, pose, and grasp identification for fractured objects. We also share our source code enabling replication of our fracturing process to other categories of object models in order to advance research involving identification, repair, and assembly of fractured objects. The dataset is shared on HuggingFace at: https://huggingface.co/datasets/tars-home/GeometricBreaks.