Sym3DNet: Symmetric 3D Prior Network for Single-View 3D Reconstruction

Area of Science:

• Computer Vision
• 3D Object Reconstruction
• Artificial Intelligence

Background:

• 3D object reconstruction is vital for tasks like recognition and analysis.
• Symmetry priors are crucial for handling occluded or partially observed objects.
• Existing methods struggle with recovering unseen parts of 3D objects.

Purpose of the Study:

• To introduce Sym3DNet, a novel method for single-view 3D reconstruction.
• To utilize a three-dimensional reflection symmetry structure prior for enhanced reconstruction.
• To improve the accuracy and efficiency of 3D object reconstruction from single images.

Main Methods:

• Sym3DNet employs 2D-to-3D encoder-decoder networks.
• A symmetry fusion step integrates flipped and overlapped 3D shapes.
• Multi-level perceptual loss is calculated in different feature spaces for voxel-wise and global symmetry assessment.

Main Results:

• Sym3DNet demonstrates superior performance on both synthetic (ShapeNet) and real-world (Pix3D) datasets.
• The method achieves higher efficiency and accuracy compared to state-of-the-art approaches.
• Promising reconstruction results were observed even with unseen object categories, indicating strong generalization.

Conclusions:

• Sym3DNet effectively incorporates 3D reflection symmetry for robust single-view 3D reconstruction.
• The proposed multi-level perceptual loss accurately captures both local and global object symmetry.
• Sym3DNet offers a significant advancement in reconstructing 3D objects, especially under challenging conditions.