InOR-Net: Incremental 3-D Object Recognition Network for Point Cloud Representation.

Area of Science:

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

Background:

• Existing 3-D object recognition models struggle with dynamic, evolving real-world categories.
• Catastrophic forgetting of previously learned classes hinders continuous learning in 3-D object recognition.
• Identifying essential 3-D geometric features for mitigating forgetting remains a challenge.

Purpose of the Study:

• To develop an Incremental 3-D Object Recognition Network (InOR-Net) capable of continuous learning.
• To address the issue of catastrophic forgetting in 3-D object recognition models.
• To identify and leverage crucial 3-D geometric characteristics for robust incremental learning.

Main Methods:

• Introduced a novel Incremental 3-D Object Recognition Network (InOR-Net).
• Developed category-guided geometric reasoning to leverage intrinsic class information for local structure analysis.
• Proposed a critic-induced geometric attention mechanism to select beneficial 3-D geometric features.
• Implemented a dual adaptive fairness compensation strategy to address class imbalance.

Main Results:

• InOR-Net demonstrates state-of-the-art performance on public point cloud datasets.
• The model successfully recognizes new 3-D object classes without significant degradation of old class performance.
• Category-guided reasoning and attention mechanisms effectively mitigate catastrophic forgetting.

Conclusions:

• InOR-Net provides a robust solution for incremental 3-D object recognition in dynamic environments.
• The proposed methods enhance the model's ability to learn new classes while retaining knowledge of old classes.
• This work advances the field of 3-D object recognition by addressing the critical challenge of catastrophic forgetting.