HAFE-Net: A Hierarchical Adaptive Feature Enhanced Framework for Few-Shot Point Cloud Segmentation

Abstract—Point cloud segmentation faces significant challenges due to insufficient annotated samples, limiting its practical deployment in real-world scenarios. To address this, we introduce few-shot learning to enhance model generalization capabilities under limited data conditions. This paper proposes a Hierarchical Adaptive Feature Enhancement Network (HAFE-Net), which establishes a robust feature learning framework through hierarchical collaboration between parametric and non-parametric encoders. The framework integrates two novel components: 1) an Adaptive Feature Enhancement (AFE) module employing randomized quantization to diversify intermediate features, effectively addressing representation homogeneity in few-shot scenarios; and 2) an Adaptive Query Enhancement (AQE) module leveraging cross-attention mechanisms to propagate prototype information into query features, significantly improving segmentation accuracy in complex environments. Extensive experiments on S3DIS and ScanNet datasets demonstrate HAFE-Net’s superiority, achieving improvements of up to 9.14% and 31.08% respectively. Notably, the single-metric enhancement on ScanNet exceeds the 30% performance threshold, conclusively verifying the method’s cross-scenario generalization capabilities.