Deep Spatiality: Unsupervised Learning of Spatially-Enhanced Global and Local 3D Features by Deep Neural Network with...

Area of Science:

• Computer Vision
• 3D Shape Analysis
• Machine Learning

Background:

• Bag-of-Words representations improve 3D shape analysis.
• Encoding spatial relationships enhances discriminability.
• Existing methods face challenges with mesh resolution, topology, orientation, and transformations.

Purpose of the Study:

• To propose a novel unsupervised spatial learning framework, Deep Spatiality (DS), for 3D shape analysis.
• To address limitations in encoding spatial relationships for improved shape discriminability.
• To develop a deep neural network approach for robust feature learning.

Main Methods:

• DS utilizes a spatial context extractor to capture local spatial relationships.
• A directed circular graph and relative spatial matrix encode pairwise virtual word positions.
• Singular Value Decomposition (SVD) generates raw spatial representations for deep context learning.
• A deep context learner with a coupled softmax layer learns global and local shape features.

Main Results:

• DS effectively encodes spatial relationships among virtual words on 3D shapes.
• The framework demonstrates unsupervised learning of both local and global shape features.
• Experimental results show DS outperforms existing state-of-the-art methods in 3D shape analysis.

Conclusions:

• Deep Spatiality (DS) offers a robust solution for unsupervised spatial learning in 3D shape analysis.
• The proposed framework successfully addresses key challenges in encoding spatial information.
• DS significantly enhances the discriminability of 3D shape representations.