Balanced Residual Distillation Learning for 3D Point Cloud Class-Incremental Semantic Segmentation

2024-08-02

arXiv.org

AI Summary

This work addresses catastrophic forgetting and class imbalance in continual incremental learning (CIL) for 3D point cloud semantic segmentation. To mitigate these challenges, we propose a novel framework featuring two core contributions: (1) a residual distillation strategy that preserves knowledge from previous tasks via dynamic network expansion and historical feature refinement; and (2) a balanced pseudo-labeling mechanism that jointly optimizes discriminative capability for both old and new classes through point-wise feature alignment and a class-balanced loss. Evaluated on standard point cloud CIL benchmarks, our method achieves significant improvements: +5.2% in mean Intersection-over-Union (mIoU), a 37% reduction in backward transfer degradation (i.e., forgetting rate on old classes), and establishes new state-of-the-art performance—particularly under severe class imbalance scenarios.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Computer Vision: Segmentation
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Class-incremental learning (CIL) enables continuous learning of new classes while mitigating catastrophic forgetting of old ones. For the performance breakthrough of CIL, it is essential yet challenging to effectively refine past knowledge from the base model and balance it with new learning. However, such a challenge has not been considered in current research. This work proposes a balanced residual distillation learning framework (BRDL) to address this gap and advance CIL performance. BRDL introduces a residual distillation strategy to dynamically refine past knowledge by expanding the network structure and a balanced pseudo-label learning strategy to mitigate class bias and balance learning between old and new classes. We apply the proposed BRDL to a challenging 3D point cloud semantic segmentation task where the data is unordered and unstructured. Extensive experimental results demonstrate that BRDL sets a new benchmark with an outstanding balance capability in class-biased scenarios.

Problem

• Class Incremental Learning
• 3D Point Clouds
• Knowledge Balance

Innovation

• BRDL
• Category Incremental Learning
• 3D Point Clouds