Parameterize Structure with Differentiable Template for 3D Shape Generation

TL;DR

This paper proposes the method that parameterizes the shared structure in the same category using a differentiable template and corresponding fixed-length parameters, and utilizes the boundaries of three-view renderings of each cuboid to further describe the inside details.

Abstract

Structural representation is crucial for reconstructing and generating editable 3D shapes with part semantics. Recent 3D shape generation works employ complicated networks and structure definitions relying on hierarchical annotations and pay less attention to the details inside parts. In this paper, we propose the method that parameterizes the shared structure in the same category using a differentiable template and corresponding fixed-length parameters. Specific parameters are fed into the template to calculate cuboids that indicate a concrete shape. We utilize the boundaries of three-view drawings of each cuboid to further describe the inside details. Shapes are represented with the parameters and three-view details inside cuboids, from which the SDF can be calculated to recover the object. Benefiting from our fixed-length parameters and three-view details, our networks for reconstruction and generation are simple and effective to learn the latent space. Our method can reconstruct or generate diverse shapes with complicated details, and interpolate them smoothly. Extensive evaluations demonstrate the superiority of our method on reconstruction from point cloud, generation, and interpolation.

Figures (9)

• Figure 1: We utilize parameters to control the position of cuboids through a differentiable template. Each cuboid is defined as a "stick" with two control points and sizes. The object is represented by the combination of cuboids. We employ three-view boundaries (black contours) to represent the details inside each cuboid. Here, we only store the red vertexes.
• Figure 2: Several relationships are utilized in differentiable templates including joint (1)(2), restriction (4)(5), line (6), and symmetry (3)(7), in the 2D version. An example (Chair-4) is shown at the bottom right to illustrate the utilization of these relationships.
• Figure 3: Reconstruction results from point clouds of different methods on Chair category. Here, we display the reconstruction results of a surface reconstruction method (PGR) for reference to evaluate the quality of our reconstructed shapes with details.
• Figure 4: An optimization example on Table-4 category of our method and ShapeAssembly. Each shape is optimized 1000 iterations.
• Figure 5: The generated shapes of our method, StructreNet, ShapeAssembly, SALAD, and DSG-Net on Chair and Table categories.