Hierarchical Spatial and Channel Aggregation for Cross-domain Few-shot Segmentation

Cross-domain Few-shot Segmentation (CD-FSS) aims to enable models trained on a source domain to accurately segment novel classes in a target domain with minimal labeled data. A primary challenge in this task is that models often suffer from semantic and attribute over-alignment, where they become overly specialized to the specific semantic granularities and attribute distributions of the source domain, leading to poor generalization when target domains exhibit different characteristics.

The authors propose the Dual Hierarchical Aggregation Network (DHANet), which introduces three key mechanisms to improve cross-domain transfer:

1. Hierarchical Spatial Aggregation (HSA): This module uses learnable spatial slots to aggregate pixel features at multiple scales, creating hierarchical semantic-enhanced features that allow the model to adapt to varying semantic granularities.
2. Hierarchical Channel Aggregation (HCA): This module aggregates features along the channel dimension to capture diverse attribute representations, preventing the model from becoming overly dependent on specific channel-wise alignments that may not hold in the target domain.
3. Online Probabilistic Semantic Bank (OPSB): During inference, this module maintains and updates class-specific probability distributions based on high-confidence query predictions. By sampling pseudo-prototypes from these distributions, the model effectively augments the limited support information available for novel classes.

Extensive experiments across four target-domain datasets demonstrate that DHANet achieves state-of-the-art performance in CD-FSS tasks. By simultaneously addressing semantic and attribute over-alignment, the model exhibits superior robustness to domain shifts compared to existing methods that focus primarily on global style alignment or single-granularity feature decoupling.