Robust3D-CIL: Robust Class-Incremental Learning for 3D Perception

2025-03-18

AI Summary

To address severe catastrophic forgetting in 3D class-incremental learning (CIL) caused by continuously emerging categories and unknown corruptions in real-world scenarios (e.g., autonomous driving, robotics), this paper proposes a robust 3D CIL framework. The method introduces two key innovations: (1) a diversity-aware exemplar selection strategy based on farthest point sampling to enhance buffer representativeness; and (2) a lightweight point cloud downsampling replay mechanism that mitigates corruption-induced forgetting under memory constraints—without requiring full model retraining. The framework is compatible with mainstream 3D detection and segmentation backbones. Evaluated on multiple 3D CIL benchmarks, our approach outperforms existing replay-based methods by 2–11% on average, demonstrating significantly improved generalization against corrupted data and enhanced stability in continual learning.

Technology Category

• Machine Learning: Life-Long and Continual Learning
• Computer Vision: 3D Computer Vision
• Intelligent Robots: Learning & Optimization for ROB

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

3D perception plays a crucial role in real-world applications such as autonomous driving, robotics, and AR/VR. In practical scenarios, 3D perception models must continuously adapt to new data and emerging object categories, but retraining from scratch incurs prohibitive costs. Therefore, adopting class-incremental learning (CIL) becomes particularly essential. However, real-world 3D point cloud data often include corrupted samples, which poses significant challenges for existing CIL methods and leads to more severe forgetting on corrupted data. To address these challenges, we consider the scenario in which a CIL model can be updated using point clouds with unknown corruption to better simulate real-world conditions. Inspired by Farthest Point Sampling, we propose a novel exemplar selection strategy that effectively preserves intra-class diversity when selecting replay exemplars, mitigating forgetting induced by data corruption. Furthermore, we introduce a point cloud downsampling-based replay method to utilize the limited replay buffer memory more efficiently, thereby further enhancing the model's continual learning ability. Extensive experiments demonstrate that our method improves the performance of replay-based CIL baselines by 2% to 11%, proving its effectiveness and promising potential for real-world 3D applications.

Problem

• Adapting 3D perception models to new data and categories efficiently.
• Handling corrupted 3D point cloud data in class-incremental learning.
• Enhancing continual learning ability with limited replay buffer memory.

Innovation

• Novel exemplar selection strategy preserves intra-class diversity
• Point cloud downsampling-based replay method enhances memory efficiency
• Robust3D-CIL improves CIL performance by 2% to 11%