A Deep Learning Method for 3D Object Classification and Retrieval Using the Global Point Signature Plus and Deep Wide...

Area of Science:

• Computer Vision
• Deep Learning
• 3D Object Recognition

Background:

• 3D object classification and retrieval are crucial for applications like robotics and autonomous driving.
• Existing view-based methods often use multiple views, complicating network structures.
• Voxelization and Point Cloud methods are alternative approaches for 3D data processing.

Purpose of the Study:

• To propose a novel and efficient deep learning framework for 3D object classification and retrieval.
• To address the complexity issues associated with multi-view deep learning methods.
• To enhance the capture of 3D object shape information using a single view.

Main Methods:

• A novel descriptor, Global Point Signature Plus (GPSPlus), was developed to capture detailed shape information from a single 3D object view.
• 3D models were converted into colored 2D projections (32x32x3 matrices) using GPSPlus.
• A Deep Wide Residual Network (GPSP-DWRN) with a single Convolutional Neural Network (CNN) structure processed the 2D projection data.

Main Results:

• The GPSP-DWRN framework demonstrated superior performance in 3D object retrieval on the Shapnetcore55 dataset.
• The method achieved state-of-the-art results for 3D object classification on the ModelNet10 and ModelNet40 datasets.
• The proposed approach effectively handles spatial information loss with a simplified network structure.

Conclusions:

• The GPSP-DWRN framework offers a more efficient and effective solution for 3D object classification and retrieval.
• GPSPlus descriptor enhances shape information capture, leading to improved performance.
• The single CNN structure simplifies the network, reducing computational complexity compared to multi-view methods.