3D Generative Design for Non-Experts: Multiview Perceptual Similarity with Agent-Based Reinforcement Learning

3D Generative Design for Non-Experts: Multiview Perceptual Similarity with Agent-Based Reinforcement Learning

Stuart-Smith, Robert; Danahy, Patrick

Full Article:

Advances in additive manufacturing allow architectural elements to be fabricated with increasingly complex geometrical designs, however, corresponding 3D design software requires substantial knowledge and skill to operate, limiting adoption by non-experts or people with disabilities. Established non-expert approaches typically constrain geometry, topology, or character to a pre-established configuration, rather than aligning to figural and aesthetic characteristics defined by a user. A methodology is proposed that enables a user to develop multi-manifold designs from sketches or images in several 3d camera projections. An agent-based design approach responds to computer vision analysis (CVA) and Deep Reinforcement Learning (RL) to design outcomes with perceptual similarity to user input images evaluated by Structural Similarity Indexing (SSIM). Several CVA and RL ratios were explored in training models and tested on untrained images to evaluate their effectiveness. Results demonstrate a combination of CVA and RL motion behavior can produce meshes with perceptual similarity to image content.

Advances in additive manufacturing allow architectural elements to be fabricated with increasingly complex geometrical designs, however, corresponding 3D design software requires substantial knowledge and skill to operate, limiting adoption by non-experts or people with disabilities. Established non-expert approaches typically constrain geometry, topology, or character to a pre-established configuration, rather than aligning to figural and aesthetic characteristics defined by a user. A methodology is proposed that enables a user to develop multi-manifold designs from sketches or images in several 3d camera projections. An agent-based design approach responds to computer vision analysis (CVA) and Deep Reinforcement Learning (RL) to design outcomes with perceptual similarity to user input images evaluated by Structural Similarity Indexing (SSIM). Several CVA and RL ratios were explored in training models and tested on untrained images to evaluate their effectiveness. Results demonstrate a combination of CVA and RL motion behavior can produce meshes with perceptual similarity to image content.

Palavras-chave: - -

DOI: 10.5151/sigradi2022-sigradi2022_53

Referências bibliográficas

• [1]

Como citar:

Stuart-Smith, Robert; Danahy, Patrick; "3D Generative Design for Non-Experts: Multiview Perceptual Similarity with Agent-Based Reinforcement Learning", p-115-126. In: XXVI International Conference of the Iberoamerican Society of Digital Graphics. São Paulo: Blucher, 2023.
ISSN 23186968, DOI 10.5151/sigradi2022-sigradi2022_53