Referring Video Object Segmentation (R-VOS) demands precise visual comprehension and sophisticated cross-modal reasoning to segment objects in videos based on descriptions from natural language. Addressing this challenge, we introduce the Cross-modal Spectral Fusion Model (CSF). Our model incorporates a Multi-Scale Spectral Fusion Module (MSFM), which facilitates robust global interactions between the modalities, and a Consensus Fusion Module (CFM) that dynamically balances multiple prediction vectors based on text features and spectral cues for accurate mask generation. Additionally, the Dual-stream Mask Decoder (DMD) enhances the segmentation accuracy by capturing both local and global information through parallel processing. Tested on three datasets, CSF surpasses existing methods in R-VOS, proving its efficacy and potential for advanced video understanding tasks.
Huang et al. (Fri,) studied this question.