VERAM: View-Enhanced Recurrent Attention Model for 3D Shape Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Multi-view deep neural networks are effective for 3D shape classification.
• Current methods using max or average pooling lack view selection, hindering applications like robotic object recognition.
• Existing attention-based models suffer from unbalanced training between view estimation and classification subnetworks.

Purpose of the Study:

• To present VERAM, a view-enhanced recurrent attention model for accurate 3D shape classification.
• To address the unbalanced training issue in attention-based 3D shape classification models.
• To improve multi-view feature fusion through active view selection.

Main Methods:

• Developed VERAM, a recurrent attention model for sequential view selection.
• Implemented three view-enhancement strategies: improved gradient flow, informative reward function for reinforcement learning, and a novel loss function to prevent view duplication.
• Utilized grayscale images and AlexNet CNN architecture for experiments.

Main Results:

• VERAM achieved state-of-the-art performance on ModelNet10 and ModelNet40 datasets with only 9 views.
• Instance-level accuracy reached 95.5% on ModelNet10 and 93.7% on ModelNet40.
• Class-level accuracy reached 95.3% on ModelNet10 and 92.1% on ModelNet40.

Conclusions:

• VERAM demonstrates superior 3D shape classification accuracy through active view selection.
• The proposed view-enhancement strategies effectively resolve training imbalances in attention models.
• VERAM offers a promising approach for multi-view active object recognition and 3D shape analysis.