• 강사 : 성민혁 교수 (카이스트 전산학부)  
• 일시 : 2021. 4. 8 (목) 17:00~18:30

3D data arising from scanning with depth sensors or modeling by designers have a unique characteristic — it matches the actual physical form of an object as it presents in the real world. Hence, unlike 2D images containing a projected view, it directly enables understanding of how the objects are composed and structured in the physical space.

In this talk, I will first introduce some challenges remaining in 3D data processing and describe how deep learning techniques have been exploited to solve them. Then, for the specific case of understanding the compositional structure of 3D shapes, I'll present our novel methodologies based on learning from a large collection of data and demonstrate various downstream applications. The main ideas are organized into two parts: decomposition and composition. In the decomposition, I will first discuss some ideas of parsing 3D raw scanned data and decomposing them into geometric primitives. For solving the model estimation problem more robustly, I will propose an end-to-end neural network pipeline that combines the best features of supervised learning and classical optimization. Also, there exist multiple cases when the compositional structures (e.g., segmentation and keypoints) are given for each of the 3D shapes in the database but without any consistency. For such cases, I will introduce an autoencoder-style neural network that can bridge the information defined on independent shapes without the supervision of correspondences. Lastly, in the composition, I will show how a part-based representation of 3D objects, fusing a discrete and combinatorial global structure with continuous local geometry space, can facilitate creating and editing shapes by efficiently exploring the 3D object space. I will also end my talk by sharing my recent work on neural-network-based 3D content creation.