Text-Injected Discriminative Model for Remote Sensing Visual Grounding

Remote Sensing Visual Grounding (RSVG) requires fine-grained understanding of language descriptions to localize the specific image regions. Conventional methods typically employ a pipeline of separate visual and textual encoders and a fusion module. However, as visual and textual features are extracted independently, they tend to lack semantic focus on object features during extraction, leading to suboptimal object focus. While some recent attempts have incorporated textual cues into visual feature extraction, they often design complex fusion modules. To address this, we introduce a simple fusion strategy to integrate textual information into visual backbone networks with minimal architectural changes. Moreover, most of the current works use common object detection losses, which only focus on the features inside the bounding box and neglect the background features. In remote sensing images, the high visual similarity between objects can confuse models, making it difficult to locate the correct target accurately. To this end, we design a novel attention regularization strategy to enhance the model’s ability to distinguish similar features outside bounding box regions. Experiments on three benchmark datasets demonstrate the promising performance of our approach.