Progressive Deep Learning Framework for Recognizing 3D Orientations and Object Class Based on Point Cloud...

Area of Science:

• Computer Vision
• Machine Learning
• 3D Object Recognition

Background:

• Estimating 3D object orientation and class using deep learning is challenging due to the continuous nature of orientation variations.
• Existing regression or classification methods often struggle with generalization and accuracy for full three-axis orientation estimation.

Purpose of the Study:

• To present a novel progressive deep learning framework, 3D POCO Net, for accurate and efficient estimation of full 3D object orientations and classes.
• To improve upon the limitations of current deep learning approaches in handling continuous orientation variations.

Main Methods:

• Developed 3D POCO Net, a framework using four PointNet-based networks for independent object class and rotation axis representation.
• Employed association subnetworks to progressively fine-tune independent networks by mapping global features.
• Combined high-precision classification with weighted regression for accurate orientation estimation.

Main Results:

• Achieved an orientation regression error of approximately 2.5° and 90% accuracy in object classification for general three-axis orientation estimation.
• Demonstrated high efficiency in network complexity through a progressive framework.
• Showcased the utility of pre-trained 3D POCO Net for transfer learning on partial point clouds of occluded objects.

Conclusions:

• 3D POCO Net offers a highly accurate and efficient solution for 3D object orientation and classification tasks.
• The framework's progressive nature and combined classification-regression approach overcome limitations of previous methods.
• Pre-trained models can be effectively utilized for orientation and class estimation from incomplete 3D data.