The newly proposed Generalized Referring Expression Segmentation (GRES) amplifies the formulation of classic RES by involving multiple/non-target scenarios. Recent approaches focus on optimizing the last modality-fused feature which is directly utilized for segmentation and object-existence identification. However, the attempt to integrate all-grained information into a single joint representation is impractical in GRES due to the increased complexity of the spatial relationships among instances and deceptive text descriptions. Furthermore, the subsequent binary target justification across all referent scenarios fails to specify their inherent differences, leading to ambiguity in object understanding. To address the weakness, we propose a $\textbf{H}$ierarchical Semantic $\textbf{D}$ecoding with $\textbf{C}$ounting Assistance framework (HDC). It hierarchically transfers complementary modality information across granularities, and then aggregates each well-aligned semantic correspondence for multi-level decoding. Moreover, with complete semantic context modeling, we endow HDC with explicit counting capability to facilitate comprehensive object perception in multiple/single/non-target settings. Experimental results on gRefCOCO, Ref-ZOM, R-RefCOCO, and RefCOCO benchmarks demonstrate the effectiveness and rationality of HDC which outperforms the state-of-the-art GRES methods by a remarkable margin. Code will be available $\href{https://github.com/RobertLuo1/HDC}{here}$.
翻译:暂无翻译