Multi-View 3D Shape Recognition via Correspondence-Aware Deep Learning.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Shape Analysis

Background:

• Multi-view learning is crucial for 3D shape recognition using 2D images.
• Existing methods underutilize spatial and symmetry cues from view correspondences.
• Explicitly leveraging correspondences can enhance geometric feature extraction.

Purpose of the Study:

• To propose a novel correspondence-aware representation (CAR) module.
• To explicitly exploit intra-view and cross-view correspondences for 3D shape recognition.
• To develop an effective deep learning model, CAR-Net, for 3D shape classification and retrieval.

Main Methods:

• Developed a CAR module to find correspondences via kNN search in semantic space.
• Aggregated shape features from correspondences using learned transformations.
• Incorporated spatial relations of correspondences (viewpoint, intra-view location).
• Integrated the CAR module with a ResNet-18 backbone to create CAR-Net.

Main Results:

• The CAR module effectively captures geometric cues from correspondences.
• CAR-Net demonstrates superior performance in 3D shape classification and retrieval tasks.
• Experiments validate the effectiveness of the proposed CAR module and CAR-Net.

Conclusions:

• Explicitly modeling view correspondences significantly improves 3D shape recognition.
• The CAR module provides a powerful new approach for feature representation in multi-view learning.
• CAR-Net offers a state-of-the-art solution for 3D shape classification and retrieval.