This paper presents the first learning-based generative pipeline for effectively creating 3D LEGO® 1 models. This task is very challenging due to the lack of dedicated representations and datasets for learning coherently-connected bricks arrangements, as well as an immense design space that is combinatorial in nature. We approach this task by focusing on creating LEGO® micro buildings. Our contributions are four-fold. First, we propose the LEGO® semantic volume representation to encode LEGO® models, considering the bricks types and bricks connections, while allowing back-propagation learning. Second, we further consider the transformative nature of LEGO ® to atomize the semantic volume and formulate a generative model to learn the representation. Third, we build a rich dataset of micro buildings for model learning. Last, we design the progressive reconstructor to create 3D LEGO® models from the generated representations, while ensuring bricks connections. We employed our pipeline to create LEGO ® micro buildings with a wide array of bricks types, demonstrating its strong capability of learning diverse micro-building styles and producing assemble-able LEGO® models. Further, we performed various quantitative evaluations, ablations, and a user study to show the compelling capability of our approach in terms of generative quality, fidelity, and diversity.
Read full abstract