Introduction
Standard Few-Shot Class-Incremental Learning (FSCIL) aims to help models learn new categories from limited examples while retaining knowledge of base classes. However, real-world applications often face a significant domain shift: base classes are typically clean, synthetic 3D models, whereas incremental classes are noisy, occluded real-world scans. Existing methods often struggle with this shift, either forgetting base knowledge or failing to adapt to the new, noisy domain. This paper introduces a novel framework for Cross-Domain FSCIL in 3D point-cloud recognition, treating new objects as "out-of-distribution" (OOD) relative to a synthetic base model.
Methodology
The proposed approach utilizes a three-pronged strategy to address domain shift and catastrophic forgetting:
- OOD-Based Task Prediction: The system first acts as an OOD detector. By determining if a sample resembles the synthetic base data, it routes the input to either the base classifier or a domain-specific module for new classes.
- Scaled Cosine Classifier: Unlike standard linear classifiers with unbounded logits, the authors employ a scaled cosine classifier. Its bounded similarity scores provide a more reliable confidence signal for OOD detection.
- Two-Branch Architecture & Dual Replay: The model maintains a frozen base-class branch to prevent forgetting. A separate branch is fine-tuned for new classes. To preserve knowledge across multiple incremental sessions, a "Dual Replay" strategy is used, storing both physical examples (hard labels) and previous output scores (soft labels).
Experimental Setup
The study evaluates synthetic-to-real transitions across several benchmarks:
- M2S: ModelNet → ScanObjectNN
- S2C: ShapeNet → CO3D
- S2S: ShapeNet → ScanObjectNN
These benchmarks test the model's ability to generalize from clean synthetic data to noisy, real-world 3D scans.
Results and Discussion
The framework was evaluated using Harmonic Accuracy ($A_h$) to ensure a balanced assessment of base and new class performance. On the M2S benchmark, the method achieved a harmonic accuracy of 54.3%, significantly outperforming state-of-the-art FSCIL methods like FACT (11.9%) and ALICE (16.0%).
Key insights include:
- Incremental Learning as an OOD Problem: Decoupling the identification of new classes from their classification prevents the confusion that typically leads to forgetting.
- Logits Over Probabilities: Maximum logits from the cosine classifier were found to be superior to Softmax probabilities for OOD detection, as they better reflect model uncertainty.
- Fine-tuning Depth: Freezing shallow layers while fine-tuning deeper, class-specific layers provides the optimal balance between stability and adaptability.
Conclusion
While the OOD detection strategy is highly effective, its reliance on a threshold ($\gamma$) may be challenging when domain shifts are minimal. Future work could explore generative replay to eliminate the need for a physical data buffer. Nevertheless, this framework sets a new standard for cross-domain incremental learning in 3D point-cloud recognition, achieving nearly 45% accuracy on new classes in final sessions—a 4-5x improvement over previous methods.
Summary updated