Geometric Breaks: A Dataset of Synthetically Generated Fractured Objects

The Geometric Breaks dataset consists of 3D models of 25,249 broken object models and their repair parts. The broken models are generated by subtracting fracturing shapes from a total of 22,165 3D complete object models. Complete object models have been acquired from the Google Scanned Object (GSO) dataset (https:\/\/app.gazebosim.org\/GoogleResearch\/fuel\/collections\/Scanned%20Objects%20by%20Google%20Research), and 8 classes in the ShapeNetCore.v2 dataset (https:\/\/huggingface.co\/datasets\/ShapeNet\/ShapeNetCore). To generate the dataset, we have waterproofed complete meshes, normalized them to a [-0.5,0.5]^3 cube, and generated broken shapes and repair counteparts by conducting Boolean subtraction with a fracturing shape from the waterproofed complete mesh. Fracturing shapes come from one of three geometric primitives: cube, icosphere, and subivided icosphere, randomly translated and rotated to generate breaks over various parts of the original model. Vertices on the primitive surface are randomly perturbed to simulate roughness. Since the subtraction process generates multiple fragments, the dataset also contains individual 3D models of the fragmented parts. Models from following ShapeNetCore.v2 classes are used: Jar (03593526), Bottle (02876657), Mug (03797390), Airplane (02691156), Chair (03001627), Car (02958343), Sofa (04256520), Table (04379243).