2026· IEEE Geoscience and Remote Sensing Letters· Vol 23, pp. 6023105-6023105· 0 citations· 20 references
TL;DR
A category-to-instance framework built on Grounding DINO (C2I-DINO), which introduces category-focused prompting (CFP), which retrieves learnable category-specific suffix vectors and appends them to the projected text memory before cross-modal fusion, and spatial instance learning (SIL), which uses ranking and suppression losses to separate the matched query from high-response queries at incorrect locations during training.
Abstract
Remote sensing visual grounding (RSVG) aims to localize objects described by natural-language expressions in large-scale aerial images. Compared with conventional visual grounding, RSVG contains many repeated objects and long-range spatial relations, requiring a model to preserve the requested category cue while distinguishing the referred instance from similar objects. To address this challenge, we propose a category-to-instance framework built on Grounding DINO (C2I-DINO). It introduces category-focused prompting (CFP), which retrieves learnable category-specific suffix vectors and appends them to the projected text memory before cross-modal fusion, and spatial instance learning (SIL), which uses ranking and suppression losses to separate the matched query from high-response queries at incorrect locations during training. CFP strengthens the category cue, while SIL provides training-only supervision. Extensive experiments on RSVG, OPT-RSVG, and DIOR-RSVG demonstrate that C2I-DINO achieves the best reported results on most metrics under the evaluated protocols. Code is available at https://github.com/YongpengWen/C2I-DINO.git
Visual grounding in remote sensing images aims to locate objects described by referring expressions. Most existing methods predict horizontal bounding boxes, which are often inaccurate for objects with arbitrary orientations. To address this limitation, we introduce O$^2$-VG, a family of models for oriented object visu...
Remote sensing image captioning (RSIC) plays a significant role in environmental surveillance and disaster evaluation. However, existing methods often struggle with the complex semantics of remote sensing imagery and lack sufficient domain-specific prior knowledge, leading to generic descriptions and semantic deviation...
Chang-Zhi Wang, S. Jiao, Yu-Han Zhu et al.· IEEE Geoscience and Remote S...· 0 citations
Source-only domain generalization (DG) for remote sensing semantic segmentation seeks reliable predictions in unseen domains without target-domain data during training. Although vision foundation models offer transferable knowledge, category recognition remains sensitive to source-specific visual cues, while task-speci...
Zi-Tong Wang, Shuchang Lyu, Meng Li et al.· 0 citations
Open-prompt detectors allow users to specify targets with text, visual exemplars, or both. We argue that existing designs underuse the visual modality in two ways. First, multiple exemplars are commonly compressed into a single class-level embedding. This textualizes visual prompting: the resulting vector plays the rol...
Jin-Long Hu, Yi Zhang, Zhi-Qiong Xia et al.· arXiv.org· 0 citations
DLPANet is proposed, a novel dual-level prototype alignment network centered on Prototype-Guided Spatial Attention, enabling simultaneous modeling of scene context and fine-grained details and demonstrates that the decoupled dual cross-attention mechanism provides superior prototype-query alignment compared to prior gl...
Mustafa Alawadi, M. Fateh· Jordanian Journal of Compute...· 0 citations
Multimodal large language models (MLLMs) must balance local detail against scene context when interpreting ultra-high-resolution (UHR) remote sensing (RS) imagery within a limited visual-input budget. Existing selection-based methods either prune tokens and select patches through relevance scoring, or crop actively thr...
Yao Zhang, Peng-Yu Dai, Wei Guo et al.· arXiv.org· 2 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.