MSDCNN: A multiscale dilated convolution neural network for fine-grained 3D shape classification.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Multi-view deep neural networks excel at 3D shape classification.
• Global features often miss fine-grained details and spatial relationships, hindering subcategory identification.
• Existing methods struggle with subtle variations among 3D shape subcategories.

Purpose of the Study:

• To propose a novel multiscale dilated convolution neural network (MSDCNN) for multi-view fine-grained 3D shape classification.
• To enhance the extraction of contextual information and spatial relationships from multi-view 3D shape data.
• To improve the accuracy of classifying 3D shapes with small inter-class variances.

Main Methods:

• A sequential view capturing module renders 12 views of the 3D shape.
• ResNeXt50 extracts semantic features per view, aggregated into a global mixed feature map.
• An attention dilated module (ADM) with attention dilated blocks (ADB) enhances context using dilated convolutions and attention mechanisms.
• A prediction module with label smoothing classifies the enhanced features.

Main Results:

• The proposed MSDCNN framework was experimentally validated on ModelNet10, ModelNet40, and FG3D datasets.
• Experimental results demonstrate the effectiveness of the MSDCNN for fine-grained 3D shape classification.
• The method shows superiority in identifying subtle variations crucial for accurate classification.

Conclusions:

• The developed MSDCNN framework significantly improves multi-view fine-grained 3D shape classification.
• The integration of multiscale dilated convolutions and attention mechanisms enhances feature representation.
• The proposed approach offers a superior solution for challenging 3D shape classification tasks.