Second-order Spectral Transform Block for 3D Shape Classification and Retrieval.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Shape Analysis

Background:

• Traditional pooling methods in deep learning may not fully capture complex geometric information in 3D shapes.
• Existing 3D shape analysis techniques often struggle with discriminative feature aggregation.

Purpose of the Study:

• To propose a novel network block, the second-order spectral transform block, for improved 3D shape retrieval and classification.
• To generalize second-order pooling to 3D surfaces with a learnable non-linear spectral transform.

Main Methods:

• Designed second-order average (SO-Avr) and max-pooling (SOMax) operations for 3D surfaces.
• Developed a learnable spectral transform using a mixture of power functions for non-linear feature mapping.
• Integrated the proposed block into existing network architectures for 3D shape analysis.

Main Results:

• Achieved a 7.2% improvement in first-tier retrieval accuracy on the SHREC’14 Real dataset.
• Attained state-of-the-art classification accuracy on the ModelNet40 dataset.
• Demonstrated superior performance in 2D image classification compared to traditional second-order pooling methods.

Conclusions:

• The proposed second-order spectral transform block effectively enhances 3D shape retrieval and classification performance.
• The block's flexibility allows integration into various network architectures, boosting their capabilities.
• Theoretical and experimental analyses confirm the stability of the proposed method.