Intra-class complex object shape representation towards high resolution

Abstract

An intra-class complex object shape representation architecture (IConv-DAE) is proposed. It outperforms prior work in 3D shape completion and reconstruction through data-driven learning in forms of volumetric representation. The main mark of this architecture is the improved performance for a more complex intra-class object shape representation. The shape representation has lots of complex shape variants and improved resolution of volumetric representation from 30 × 30 × 30 up to 100 × 100 × 100. In our experiments, the designed architectures are applied for testing generative ability of our proposed architecture for completed shape, noised shape, slice-missing shape and structure-missing shape. And the improved performance over existing deep neural network architectures can be achieved.