L3DOC: Lifelong 3D Object Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• 3D object classification is crucial but current methods struggle with new tasks and lifelong learning due to complex 3D data.
• Existing lifelong learning models often suffer from catastrophic forgetting when processing irregular 3D geometry data.

Purpose of the Study:

• To propose a novel Lifelong 3D Object Classification (L3DOC) model capable of continuously learning new 3D object classification tasks.
• To address the challenges of catastrophic forgetting and performance degradation in lifelong learning scenarios with 3D data.

Main Methods:

• Introduced a point-knowledge factorization architecture to capture and store cross-task common knowledge within a 3D neural network.
• Employed a task-relevant knowledge distillation mechanism, including point-knowledge and transforming-space distillation, to connect new tasks with previous ones.

Main Results:

• The L3DOC model effectively learns new 3D object classification tasks sequentially without compromising previously learned knowledge.
• Experimental results on point cloud benchmarks demonstrate the superiority of L3DOC over existing state-of-the-art lifelong learning methods.

Conclusions:

• L3DOC represents the first deep learning approach for lifelong 3D object classification, mimicking human learning.
• The proposed model offers a robust solution for continuous learning in 3D object classification, overcoming catastrophic forgetting.