CRS-Det: Consistent Region Semantic Modeling for Remote Sensing Open-Vocabulary Object Detection
Abstract
Remote sensing open-vocabulary object detection (RSOVD) aims to detect objects beyond the categories annotated during training. Although vision–language models (VLMs) have significantly advanced this task, existing approaches often suffer from region-level semantic instability, where region representations are highly sensitive to contextual variation and background interference, limiting generalization to novel categories. To address this issue, we propose CRS-Det, which introduces feature-level region semantic modeling to stabilize region representations rather than relying on image-level region discrimination. In particular, we propose multiview region semantic consistency (MRSC) to learn stable region representations by enforcing semantic consistency across multiple views. To reduce the influence of irrelevant background regions, we design region-selective aggregation (RSA), which learns activation-guided spatial weights to emphasize object-relevant visual responses and suppress irrelevant background features. In addition, we introduce prototype-guided semantic alignment (PGSA) to refine textual embeddings using learnable semantic prototypes, improving region–text alignment without introducing image-conditioned prompts. Experiments on four remote sensing benchmarks demonstrate consistent improvements over the controlled baseline and competitive performance against representative open-vocabulary detectors, with particularly notable gains on novel categories.