ResearchPod Summary
Visual grounding in remote sensing aims to locate objects based on natural language descriptions. Traditional methods typically predict horizontal bounding boxes, which are often inaccurate for remote sensing images where objects exhibit arbitrary orientations, dense distributions, and significant scale variations. This paper introduces the O2-VG framework, a family of models designed to address these challenges through three complementary approaches: O2-VG-Trans (a discriminative transformer), O2-VG-Uni (a universal proposal generator), and O2-VG-VLM (a generative vision-language model).
To support training and evaluation, the authors constructed the DIOR-R-RSVG dataset. By integrating data from existing remote sensing benchmarks (DIOR, DIOR-R, and DIOR-RSVG), the authors created a high-quality dataset containing 17,402 images and 38,320 triplets of images, referring expressions, and oriented bounding boxes. This dataset provides a more comprehensive benchmark for oriented object grounding than previous datasets, featuring larger image sizes and a wider variety of object categories.
This work provides a flexible, unified solution for oriented object visual grounding, bridging the gap between discriminative transformers and generative vision-language models. By explicitly modeling object orientation and providing a large-scale oriented grounding dataset, the O2-VG family enables more precise localization in complex geographic scenes, which is critical for applications like urban planning, disaster monitoring, and defense.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.