STD-Net: Structure-Preserving and Topology-Adaptive Deformation Network for Single-View 3D Reconstruction

Area of Science:

• Computer Vision
• Computer Graphics
• Artificial Intelligence

Background:

• 3D reconstruction from single images is a challenging problem, with existing point cloud and volumetric methods struggling with complex topologies and fine details.
• Recent advancements in deep shape representations offer a promising direction for learning intricate 3D structures and details.

Purpose of the Study:

• To propose a novel approach, STD-Net, for accurate 3D model reconstruction from single-view images.
• To leverage mesh representations for effectively characterizing complex object structures and geometric details.

Main Methods:

• STD-Net employs an auto-encoder for object structure recovery from single images using bounding box representations.
• A topology-adaptive Graph Convolutional Network (GCN) refines vertex positions for complex mesh topologies.
• A unified mesh deformation block transforms structural boxes into detailed, structure-aware meshes.

Main Results:

• STD-Net demonstrates superior performance in reconstructing complex structures and fine geometric details compared to state-of-the-art methods.
• Evaluations conducted on ShapeNet and PartNet datasets validate the effectiveness of the proposed approach.

Conclusions:

• STD-Net offers a significant advancement in single-view 3D reconstruction, particularly for objects with intricate geometry.
• The mesh-based representation and proposed network architecture effectively address limitations of previous methods.