Semantics-Aware Dynamic Localization and Refinement for Referring Image Segmentation

Referring image segmentation segments an from a language expression. With the aim of producing high-quality masks, existing methods often adopt iterative learning approaches that rely on RNNs or stacked attention layers to refine vision-language features. Despite their complexity, RNN-based are subject specific encoder choices, while attention-based offer limited gains. In this work, we introduce simple yet effective alternative for progressively discriminative multi-modal The core idea our approach is leverage continuously updated query as representation target object and at each iteration, strengthen features strongly correlated weakening less related ones. As initialized by successively features, algorithm gradually shifts being localization-centric segmentation-centric. This strategy enables incremental recovery missing parts and/or removal extraneous through iteration. Compared its counterparts, method more versatile—it can be plugged into prior arts straightforwardly consistently bring improvements. Experimental results challenging datasets RefCOCO, RefCOCO+, G-Ref demonstrate advantage with respect state-of-the-art methods.