COSMOS: Context-preserving satellite memory-optimized segmentation system for ultra large-scale remote sensing images

This is the authors' abstract. We don't add key points for this paper.

Semantic segmentation of ultra-large remote sensing images faces substantial memory challenges, with virtually no previous work successfully addressing the preservation of global context in such large-scale imagery. Despite the critical importance of maintaining long-range relationships, GPU memory limitations have forced researchers to process these massive images as small disconnected patches. Our novel framework represents a breakthrough, being the first approach (with only one prior incomplete attempt in literature) to efficiently process large 4K patches on a single 24 GB GPU while preserving crucial large-scale image context. Unlike conventional tiling methods, our approach maintains global context through an innovative two-stage process: first extracting features from smaller 1K patches using encoders from powerful segmentation models (SegFormer and the Segment Anything Model), then assembling these embeddings to train a 4K decoder. This strategic separation dramatically reduces memory requirements while maintaining long-range dependencies. Through extensive experiments and ablation studies on the FBP, GID and Inria benchmark datasets, our approach achieves unprecedented performance, significantly outperforming all existing methods in both accuracy and memory efficiency.