Generating implicit object fragment datasets for machine learning

• Autores: López, Alfonso; Rueda, José Antonio; Segura, Rafael J.; Ogayar, Carlos J.; Navarro, Jose Pablo; Fuertes, José M.
• Año de publicación: 2024
• Idioma: inglés
• Tipo de recurso: artículo
• Materia: VOXEL; FRAGMENTATION; FRACTURE DATASET; VORONOI; GPU PROGRAMMING
• Institución: Consejo Nacional de Investigaciones Científicas y Técnicas

One of the primary challenges inherent in utilizing deep learning models is the scarcity and accessibility hurdles associated with acquiring datasets of sufficient size to facilitate effective training of these networks. This is particularly significant in object detection, shape completion, and fracture assembly. Instead of scanning a large number of real-world fragments, it is possible to generate massive datasets with synthetic pieces. However, realistic fragmentation is computationally intensive in the preparation (e.g., pre-factured models) and generation. Otherwise, simpler algorithms such as Voronoi diagrams provide faster processing speeds at the expense of compromising realism. In this context, it is required to balance computational efficiency and realism. This paper introduces a GPU-based framework for the massive generation of voxelized fragments derived from high-resolution 3D models, specifically prepared for their utilization as training sets for machine learning models. This rapid pipeline enables controlling how many pieces are produced, their dispersion and the appearance of subtle effects such as erosion. We have tested our pipeline with an archaeological dataset, producing more than 1M fragmented pieces from 1,052 Iberian vessels (). Although this work primarily intends to provide pieces as implicit data represented by voxels, triangle meshes and point clouds can also be inferred from the initial implicit representation. To underscore the unparalleled benefits of CPU and GPU acceleration in generating vast datasets, we compared against a realistic fragment generator that highlights the potential of our approach, both in terms of applicability and processing time. We also demonstrate the synergies between our pipeline and realistic simulators, which frequently cannot select the number and size of resulting pieces. To this end, a deep learning model was trained over realistic fragments and our dataset, showcasing similar results.
Fil: López, Alfonso. Universidad de Jaén; España
Fil: Rueda, José Antonio. Universidad de Jaén; España
Fil: Segura, Rafael J.. Universidad de Jaén; España
Fil: Ogayar, Carlos J.. Universidad de Jaén; España
Fil: Navarro, Jose Pablo. Consejo Nacional de Investigaciones Científicas y Técnicas. Centro Científico Tecnológico Conicet - Centro Nacional Patagónico. Instituto Patagónico de Ciencias Sociales y Humanas; Argentina
Fil: Fuertes, José M.. Universidad de Jaén; España