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

Researchers have made breakthroughs in Cross-domain Few-shot Segmentation (CD-FSS), achieving state-of-the-art performance on four target-domain datasets in 2026.

Two separate studies submitted in 2026 focused on improving CD-FSS, a technique that aims to learn generalizable segmentation capability from annotated samples in one domain and apply it to another with minimal annotated samples. The first study proposed the Dual Hierarchical Aggregation Network (DHANet), comprising three key modules: Hierarchical Spatial Aggregation, Hierarchical Channel Aggregation, and Online Probabilistic Semantic Bank^[1]. This method achieved state-of-the-art performance on four target-domain datasets. A second study introduced a training-free framework using a self-supervised vision encoder with three core modules: Semantic-aware Feature Re-fusion (SAFR), Adaptive Support Enhancement (ASE), and Hybrid Prototype Matching (HPM)^[2]. This framework eliminated trainable parameters to avoid training overhead and overfitting, also achieving state-of-the-art performance in CD-FSS without any training. Both studies conducted experiments on four target domain datasets, with the research being submitted in 2026^[1][2].