Survey and Evaluation of Neural 3D Shape Classification Approaches.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• 3D object classification is a key challenge in machine learning.
• Deep neural networks are widely used, but their effectiveness varies.
• Existing methods require diverse input representations and network architectures.

Purpose of the Study:

• To conduct a comprehensive survey of existing 3D object classification methods.
• To categorize these methods into a taxonomy based on common principles.
• To evaluate the performance of selected deep learning networks on 3D object classification tasks.

Main Methods:

• Extensive literature survey to identify and categorize existing 3D object classification approaches.
• Selection and evaluation of 11 deep neural networks on multiple 3D object datasets.
• Development of a framework for data conversion, training, and evaluation across different network requirements.

Main Results:

• A taxonomy of 3D object classification methods was established.
• Comparative performance analysis of 11 networks on standard and extended datasets.
• Insights into the relative strengths and weaknesses of different approaches across varying data conditions.

Conclusions:

• The study provides a structured overview and empirical evaluation of deep learning for 3D object classification.
• Performance varied across networks and datasets, highlighting the need for careful selection and adaptation.
• The released code facilitates reproducible research and future experimentation in this domain.