SeeCo: On-the-Fly Recalibration for Open-Vocabulary Remote Sensing Segmentation
Researchers propose Seeking Consensus (SeeCo), a plug-and-play framework for open-vocabulary semantic segmentation in remote sensing images. Existing methods use static inference, causing semantic ambiguity and incomplete foreground activation. SeeCo recalibrates models on-the-fly via geometric consensus learning (multi-view observations) and semantic consensus learning (textual description calibration). The framework is training-free and aims to improve performance across diverse land cover categories.
Key facts
- SeeCo is a plug-and-play framework for open-vocabulary semantic segmentation in remote sensing images.
- It addresses limitations of static inference paradigms that cause semantic ambiguity.
- Geometric consensus learning uses multi-view consistent observations.
- Semantic consensus learning adapts textual descriptions for calibration.
- The framework is training-free and recalibrates models on-the-fly.
- It targets improved performance in diverse land cover categories.
- The approach is detailed in arXiv paper 2604.26221.
- The paper is categorized as a cross-type announcement.
Entities
Institutions
- arXiv