BiRIS: A Parameter-Efficient Bidirectional Cross-Modal Alignment Framework for Referring Image Segmentation
Year: 2026
Type: conference-paper
Abstract: Parameter-Efficient Tuning (PET) provides an efficient paradigm for adapting frozen large-scale vision–language models to Referring Image Segmentation (RIS). However, RIS heavily relies on fine-grained vision–language alignment. Existing PET methods typically improve transfer performance through complex cross-modal interaction designs or additional multi-scale convolutions, while paying relatively limited attention to alignment consistency between intermediate visual and language representations, which in turn restricts the modeling of fine-grained vision–language correspondence. Motivated by this observation, we propose BiRIS, a parameter-efficient bidirectional cross-modal alignment framework for RIS. Specifically, the proposed Bidirectional Cross Aligner (BCA) constructs a compact bidirectional interaction space across multiple paired visual–language backbone layers, and its core Cross-Alignment Attention (CAA) jointly models cross-modal correlations between intermediate visual and language representations, thereby consistently improving fine-grained bidirectional alignment across layers. In addition, we introduce a Semantic Adaptive Cosine Gate (SACG), which performs fine-grained adaptive gated scaling over visual features conditioned on sentence semantics, enabling more effective modeling of the dynamic modulation of visual responses by sentence-level semantics. Extensive experiments on the RefCOCO, RefCOCO+, and G-Ref benchmarks demonstrate that BiRIS achieves competitive performance while maintaining parameter efficiency. The code is available at https://github.com/KATe-AGI/BiRIS.
Authors: Yang Li, Junqi Zhang
Institution: Tongji University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225418134
