Unsupervised Learning of 3-D Local Features From Raw Voxels Based on a Novel Permutation Voxelization Strategy.

Area of Science:

• Computer Vision
• 3-D Shape Analysis
• Machine Learning

Background:

• Hand-crafted 3-D local descriptors require significant human input.
• Unsupervised deep learning for 3-D features faces challenges with mesh irregularities and transformations.

Purpose of the Study:

• To propose an unsupervised framework for learning high-level, hierarchical 3-D local features from raw 3-D voxels.
• To overcome limitations of existing methods in handling irregular topology, mesh resolution, and orientation ambiguity.

Main Methods:

• A novel permutation voxelization strategy discretizes 3-D regions into regular voxels.
• Permutation of voxels eliminates rotation and orientation ambiguity.
• Stacked sparse autoencoders learn hierarchical patterns from permuted voxel vectors.

Main Results:

• The learned local features demonstrate superior performance in global and partial shape retrieval.
• The framework achieves state-of-the-art results in shape correspondence tasks.
• Experimental validation confirms the effectiveness of the proposed feature learning approach.

Conclusions:

• The permutation voxelization strategy is effective for unsupervised 3-D local feature learning.
• The proposed framework offers a robust solution for complex 3-D shape analysis tasks.
• This method advances the field of 3-D shape descriptors by enabling efficient and accurate feature extraction.