metadata
task_categories:
- image-segmentation
tags:
- remote-sensing
- semi-supervised-learning
Toward Stable Semi-Supervised Remote Sensing Segmentation via Co-Guidance and Co-Fusion
Project Page | Paper | GitHub
Co2S is a stable semi-supervised remote sensing (RS) segmentation framework designed to mitigate pseudo-label drift and error accumulation. It synergistically fuses priors from vision-language models (CLIP) and self-supervised models (DINOv3).
Introduction
Semi-supervised remote sensing image semantic segmentation often struggles with confirmation bias and pseudo-label drift. Co2S addresses these challenges using a heterogeneous dual-student architecture comprising two distinct ViT-based vision foundation models.
Key Features
- Heterogeneous Dual-Student Architecture: Utilizes pretrained CLIP and DINOv3 to mitigate error accumulation.
- Explicit-Implicit Semantic Co-Guidance Mechanism: Employs text embeddings and learnable queries to provide class-level guidance and enhance semantic consistency.
- Global-Local Feature Collaborative Fusion Strategy: Fuses global contextual information from CLIP with local structural details from DINOv3 for precise segmentation results.
The framework demonstrates leading performance across six popular remote sensing datasets and diverse partition protocols.