ReAttnCLIP: Training-Free Open-Vocabulary Remote Sensing Image Segmentation via Re-defined Attention in CLIP

Remote sensing image segmentation is essential for applications such as natural disaster monitoring and precision agriculture. Open-vocabulary segmentation improves flexibility by removing fixed category constraints, enabling more fine-grained scene understanding. However, unlike CLIP's pretraining objective that emphasizes global image-text alignment, segmentation requires discriminative patch-level representations for accurate pixel-wise prediction. Consequently, the quality of attention maps in the final transformer layers is critical for modeling interactions among spatial regions. Existing methods often produce suboptimal representations when capturing the complex spatial structures of remote sensing imagery. To address this issue, we refine CLIP's attention mechanism through three modifications: (1) replacing patch-to-patch attention with intermediate-layer feature similarities to better preserve spatial structure; (2) leveraging intermediate-layer attention for class-to-patch alignment to reduce classification interference; and (3) disabling the [CLS] token's selfattention to mitigate bias. Experiments on multiple remote sensing benchmarks, including building and road extraction datasets, show that our method achieves state-of-theart performance among training-free approaches.