Dual Context-Conditioned Open-Vocabulary Semantic Segmentation in Remote Sensing
IEEE Access · 2026 · 14 · 43664-43678
DOI: 10.1109/access.2026.3674757 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Sungjin Lim und 3 weiteren Autor:innen. Die Arbeit nennt 83 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: Institute of Information & communications Technology Planning & Evaluation (IITP) grants funded by the Korea governmen, Artificial Intelligence Graduate School Program, Effective Human-Machine Teaming With Multimodal Hazy Oracle Models] and by the National Research Foundation of Korea(NRF) grants funded by the Korea governmen.
Abstract (Originalsprache)
Open-vocabulary semantic segmentation (OVSS) aims to assign pixel-level labels to unseen categories during training, offering a flexible and scalable solution for real-world applications. Although vision-language models (VLMs) such as CLIP have achieved remarkable success in natural images, the extension of OVSS to remote sensing remains challenging due to heterogeneous landscapes, large-scale variations, and small objects in complex backgrounds. To preserve spatial details for improving the segmentation performance on small objects in remote sensing images, existing methods typically adopt tile-based inference with sliding windows rather than downscaling entire scenes. However, processing each tile independently without considering surrounding contexts often leads to incorrect predictions due to the absence of global scene understanding and the discontinuities at tile borders. Here, we propose training-free dual context-conditioned OVSS specifically tailored for remote sensing. Our proposed dual context conditions consist of inter-region context integration and global context integration and are incorporated into the last attention layer of the CLIP image encoder via our frequency-aware gating mechanism, enabling joint refinement of attention weights and value representations to balance between fine details and scene-level structures. Specifically, the inter-region context integration employs vision foundation models (VFMs) to measure patch-level similarities between a target tile and its neighboring tiles, providing complementary local neighboring cues and the global context integration leverages a downsampled whole image to capture structural cues via VFMs, enabling the model to incorporate global information cues. Since the reliability of local and global cues differs across scenes, our frequency-aware gating mechanism adaptively modulates their relative importance based on the image’s frequency composition. Extensive experiments on 17 benchmark datasets demonstrate that our method achieves state-of-the-art performance for training-free OVSS in remote sensing, consistently outperforming prior arts on all benchmarks.
Förderung
- Institute of Information & communications Technology Planning & Evaluation (IITP) grants funded by the Korea governmen
- Artificial Intelligence Graduate School Program
- Effective Human-Machine Teaming With Multimodal Hazy Oracle Models] and by the National Research Foundation of Korea(NRF) grants funded by the Korea governmen
Zitieren
APA
Lim, S., Kang, D. U., Cho, Y., & Chun, S. Y. (2026). Dual Context-Conditioned Open-Vocabulary Semantic Segmentation in Remote Sensing. IEEE Access, 14, 43664-43678. https://doi.org/10.1109/access.2026.3674757
Vancouver
Lim S, Kang DU, Cho Y, Chun SY. Dual Context-Conditioned Open-Vocabulary Semantic Segmentation in Remote Sensing. IEEE Access. 2026;14:43664-43678. doi:10.1109/access.2026.3674757
BibTeX
@article{Lim2026,
author = {Lim, Sungjin and Kang, Dong Un and Cho, Yeongmok and Chun, Se Young},
title = {Dual Context-Conditioned Open-Vocabulary Semantic Segmentation in Remote Sensing},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {43664--43678},
doi = {10.1109/access.2026.3674757}
}
EUVIMED