Generalized few-shot 3D point cloud segmentation requires adapting to new classes from a few annotated support samples. However, these annotations are frequently compromised by sparsity and occlusion, making them unreliable guides for model adaptation.
The paper argues that blindly adapting to such low-quality samples risks learning spurious geometric patterns rather than robust features. To address this, the authors propose an approach where uncertainty gates the gradient, enabling adaptive updates that control how much the model changes based on each support sample.
By gating updates, the method presumably down-weights unreliable samples and prevents overfitting to noise. The abstract is truncated, so details on the exact gating mechanism and experimental results are not available in the provided text.