Learning View-Based Graph Convolutional Network for Multi-View 3D Shape Analysis.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• View-based 3D shape recognition using projected 2D images is a leading approach.
• Key challenges include aggregating multi-view features and handling arbitrary object poses.
• Existing methods struggle with rotation sensitivity and effective feature fusion.

Purpose of the Study:

• To propose novel Graph Convolutional Network (GCN) models, view-GCN and view-GCN++, for enhanced 3D shape recognition.
• To address the challenges of multi-view feature aggregation and pose invariance in 3D shape recognition.
• To achieve state-of-the-art performance in 3D shape classification and retrieval.

Main Methods:

• Constructing a view-graph where multiple 2D views serve as graph nodes.
• Designing hierarchical GCNs (view-GCN) with local/non-local graph convolutions and selective view-sampling for coarsening.
• Developing view-GCN++ with local attentional graph convolution and rotation-robust view-sampling for enhanced invariance to rotation.

Main Results:

• view-GCN and view-GCN++ demonstrate superior performance on benchmark datasets (ModelNet40, ScanObjectNN, RGBD, ShapeNet Core55).
• view-GCN++ achieves invariance to transformations under the finite subgroup of the rotation group SO(3).
• Both models achieve state-of-the-art results for 3D shape classification and retrieval tasks in both aligned and rotated settings.

Conclusions:

• The proposed view-GCN and view-GCN++ models offer a powerful new approach to 3D shape recognition.
• These GCN-based methods effectively handle multi-view feature aggregation and achieve significant improvements in pose and rotation invariance.
• The models represent a substantial advancement in the field of 3D computer vision and object recognition.