cs.CVAug 27, 2026
SaveUniGeo: A Multi-modal Large Language Model for Text-Guided Cross-View Geo-Localization
Organizations: School of Computer Engineering and Science, Shanghai University, Shanghai 200444, China · Faculty of Science and Technology, and Institute of Collaborative Innovation, University of Macau, Macau 999078, China
Abstract
Text-guided drone geo-localization aims to identify a target region in a large-scale image gallery from a natural-language description. Existing methods mainly formulate this task as direct matching between an open-ended text query and candidate images. However, incomplete queries and highly similar candidates often make global cross-modal matching insufficient for reliable fine-grained localization. We propose UniGeo, a unified multimodal large language model (MLLM) for text-guided drone geo-localization. Built on a shared vision-language framework, UniGeo jointly supports geo-semantic understanding, cross-view semantic generation, and candidate-level verification. Specifically, it establishes stable correspondences among local scene elements, spatial relations, and language descriptions through geo-semantic learning, and further models semantic mappings between drone and satellite views through cross-view generation. Based on these capabilities, a plug-and-play verification module performs fine-grained discrimination among highly confusable candidates. We further introduce a multi-stage training strategy that progressively learns geo-semantic understanding, cross-view generation, and candidate verification, improving adaptation to text-guided geo-localization. Experiments demonstrate consistent improvements across multiple retrieval backbones. On GeoText-1652, UniGeo improves R@10 and mAP by 13.59 and 2.83 percentage points, respectively, validating its effectiveness for fine-grained text-guided drone geo-localization.
Figures & tables
Fig. 1 : Unified multi-task capability of UniGeo. UniGeo is a unified MLLM for text-guided geo-localization, supporting multiple geo-localization-oriented tasks such as region grounding, image captioning, region-grounded description, image-text matching, pose-aware cross-view generation, and candidate verification.
Fig. 2 : Comparison of the proposed method with existing backbones in terms of Text Query mAP and parameter numbers on GeoText-1652. We observe that our method brings consistent improvements over METER-Swin [ 16 ] , ALBEF [ 17 ] , XVLM [ 18 ] , and GeoText-1652 [ 11 ] while introducing only a marginal number of additional parameters.
Fig. 3 : The framework of UniGeo, including Progressive Learning Policy and Plug-and-Play Inference. It first converts text, image, and pose inputs into unified tokens and processes them with a shared MLLM, together with task-specific heads for text decoding, pose-aware generation, and candidate verification. The model is optimized through a progressive three-stage learning policy: Stage-I learns geo-semantic understanding with region-level grounding and spatial-relation supervision, Stage-II learns pose-aware cross-view generation under explicit geometric conditions, and Stage-III trains a hard-negative verification head for fine-grained candidate discrimination. During plug-and-play inference, UniGeo operates on the candidate pool returned by an external retrieval backbone, performs candidate-conditioned query refinement and candidate-level verification, and finally re-ranks highly confusable candidates for more reliable geo-localization.
| Method | Text Query | Image Query | ||||||
|---|---|---|---|---|---|---|---|---|
| R@1 | R@5 | R@10 | mAP | R@1 | R@5 | R@10 | mAP | |
| METER-Swin [ 16 ] | 0.98 | 2.23 | 3.07 | 1.61 | 1.43 | 3.95 | 5.81 | 2.54 |
| METER-Swin + Ours | 1.03 (+0.05) | 2.29 (+0.06) | 3.18 (+0.11) | 1.61 (+0.00) | 1.43 (+0.00) | 4.04 (+0.09) | 5.90 (+0.09) | 2.62 (+0.08) |
| ALBEF-4M [ 17 ] | 1.81 | 4.56 | 6.59 | 3.23 | 2.21 | 7.37 | 11.51 | 4.68 |
| ALBEF-4M + Ours | 1.82 (+0.01) | 5.04 (+0.48) | 7.38 (+0.79) | 3.37 (+0.14) | 3.19 (+0.98) | 9.56 (+2.19) | 13.61 (+2.10) | 6.12 (+1.44) |
| ALBEF-14M [ 17 ] | 1.84 | 4.83 | 6.69 | 3.34 | 4.59 | 13.12 | 18.95 | 8.69 |
TABLE I : Re-implemented image-text bidirectional retrieval results on GeoText-1652 [ 11 ] . The first block reports direct evaluation results of pre-trained models, the second block reports results after fine-tuning on GeoText-1652, and the last block compares recent geo-localization methods. Text Query denotes text-to-image search, and Image Query denotes image-to-text search. We report Recall@K and mAP as evaluation metrics.
| Method | Text Query | Image Query | ||||||
|---|---|---|---|---|---|---|---|---|
| R@1 | R@5 | R@10 | mAP | R@1 | R@5 | R@10 | mAP | |
| METER-Swin [ 16 ] | 2.71 | 6.73 | 8.66 | 4.44 | 3.48 | 10.32 | 14.89 | 1.04 |
| METER-Swin + Ours | 2.74 (+0.03) | 7.24 (+0.51) | 9.30 (+0.64) | 4.57 (+0.13) | 3.61 (+0.13) | 10.56 (+0.24) | 15.07 (+0.18) | 1.12 (+0.08) |
| ALBEF-4M [ 17 ] | 4.44 | 11.25 | 15.65 | 7.39 | 4.76 | 15.17 | 24.93 | 1.81 |
| ALBEF-4M + Ours | 4.50 (+0.06) | 12.68 (+1.43) | 17.48 (+1.83) | 7.89 (+0.50) | 4.84 (+0.08) | 16.92 (+1.75) | 25.71 (+0.78) | 1.97 (+0.16) |
| ALBEF-14M [ 17 ] | 5.35 | 12.91 | 17.97 | 8.72 | 5.37 | 17.51 | 26.52 | 2.13 |
TABLE II : Re-implemented image-text bidirectional retrieval results on UAVReason [ 71 ] . The first block reports direct evaluation results of pre-trained models, the second block reports results after fine-tuning on UAVReason, and the last block compares recent geo-localization methods. Text Query denotes text-to-image search, and Image Query denotes image-to-text search. We report Recall@K and mAP as evaluation metrics.
| Model | Whole-Image Caption | Region-Grounded Description | ||||||
|---|---|---|---|---|---|---|---|---|
| BLEU-1 | BLEU-2 | BLEU-4 | unigram F1 | Region-Match Acc@1 | BLEU-1 | BLEU-4 | F1 | |
| Ours | 0.42 | 0.30 | 0.16 | 0.48 | 0.71 | 0.27 | 0.06 | 0.34 |
| UniLIP [ 65 ] | 0.38 | 0.19 | 0.06 | 0.39 | 0.69 | 0.22 | 0.03 | 0.28 |
| Qwen-3.5-4B [ 67 ] | 0.29 | 0.14 | 0.03 | 0.34 | 0.70 | 0.19 | 0.02 | 0.26 |
| Bagel [ 66 ] | 0.36 | 0.23 | 0.10 | 0.42 | 0.69 | 0.18 | 0.03 | 0.26 |
TABLE III: Evaluation of query refinement and grounded description quality. We compare UniGeo with representative multimodal generation models on whole-image captioning and region-grounded description generation. Region-Match Acc@1 evaluates whether the generated description is correctly associated with the target region. Higher values indicate better performance for all metrics.
| Model | Drone Generation | Satellite Generation | |||||
|---|---|---|---|---|---|---|---|
| FID | PSNR | SSIM | MeanIdxErr | FID | PSNR | SSIM | |
| Ours | 34.45 | 15.74 | 0.18 | 15.40 | 41.34 | 13.12 | 0.16 |
| UniLIP [ 65 ] | 50.72 | 10.75 | 0.14 | 20.23 | 56.72 | 11.21 | 0.13 |
| Qwen-Image-2511 [ 73 ] | 128.04 | 10.29 | 0.16 | 19.62 | 94.58 | 10.88 | 0.13 |
| Bagel [ 66 ] | 208.18 | 7.31 | 0.05 | 21.79 | 113.55 | 7.82 | 0.05 |
TABLE IV: Evaluation of pose-aware cross-view image generation quality. We compare UniGeo with representative multimodal generation models on two cross-view synthesis directions: drone generation and satellite generation. FID, PSNR, and SSIM measure generation quality, while MeanIdxErr additionally evaluates generation accuracy in the drone generation setting. Best and second-best results are highlighted in bold and underline, respectively.
Fig. 4 : Qualitative comparison of bidirectional cross-view image generation. We compare UniGeo with other multimodal generation models on two synthesis directions. In satellite-view generation, the target image is generated under the default north-up orientation, while in drone-view generation, the viewing direction and pose are controlled by the KML metadata.
TABLE V: Ablation study and hyperparameter sensitivity. (a) evaluates the contribution of different components under the text-query and image-query settings, where CQR, CV, and HNL denote Candidate-Conditioned query refinement, candidate-level verification, and hard-negative learning, respectively. Refinement only enables CQR alone and uses the refined text query for retrieval-side ranking without CV. Verifier only enables CV alone on the baseline candidate pool without CQR. Hard Negative applies CV with HNL, while w/o Hard Negative removes HNL from the full framework. (b) reports the sensitivity to the hyperparameters and under the text-query setting. Best and second-best results are highlighted in bold and underline.
TABLE VI: Analysis of hard-negative verification learning. (a) evaluates different negative sources for training the verification head. None denotes the full framework without hard-negative verification learning, while Real HN and Syn HN denote real retrieval hard negatives and synthetic hard negatives, respectively. Avg. PairAcc denotes the average pairwise discrimination accuracy of the Stage-III verification head on held-out real and synthetic hard-negative pairs, and Gap denotes the absolute difference between the two pair accuracies. A smaller Gap indicates more balanced discrimination across different hard-negative sources. (b) studies the influence of the mixing ratio between real and synthetic hard negatives.
Fig. 5 : Effect of different settings on retrieval performance. We compare the baseline GeoText-1652 model, two variants with and , and our model. Our model consistently achieves the best R@1, R@5, and R@10 performance.
| Refinement Model | R@1 | R@5 | R@10 | mAP |
|---|---|---|---|---|
| – (Baseline) | 13.97 | 25.20 | 31.78 | 20.35 |
| InternVL3 (frozen) [ 69 ] | 14.69 | 32.01 | 42.51 | 22.64 |
| Qwen3-VL (frozen) [ 74 ] | 14.71 | 32.85 | 43.94 | 22.95 |
| Ours | 14.72 | 33.80 | 45.37 | 23.18 |
TABLE VII: Ablation Study on the refinement model under the text-query setting. The baseline retrieval model is fixed as GeoText-1652. “InternVL3” and “Qwen3-VL” denote replacing our refinement model with the off-the-shelf InternVL3-2B and Qwen3-VL-2B in the same Plug-and-Play Inference pipeline, while “Ours” denotes the proposed UniGeo refinement model.
| Method | FID | PSNR | SSIM | MeanIdxErr |
|---|---|---|---|---|
| only diff | 33.76 | 15.86 | 0.185 | 22.84 |
| w/o pose | 34.08 | 15.82 | 0.183 | 22.17 |
| w/o heading | 34.73 | 15.69 | 0.178 | 18.36 |
| 34.96 | 15.71 | 0.177 | 17.12 | |
| 36.58 | 15.49 | 0.168 | 15.86 | |
| 35.37 | 15.60 | 0.173 | 16.09 |
TABLE VIII: Ablation study of the proposed pose-aware generation objective for drone-view image generation. “only diff” denotes using only the diffusion reconstruction loss, “w/o pose” and “w/o heading” remove the pose consistency loss and heading auxiliary loss, “ ” and “ ” only change the pose loss weight to 0.1 and 0.3, “ ” only changes the heading auxiliary loss weight to 0.2.
Fig. 6 : Qualitative results of pose-aware cross-view generation. Rows 1 and 3 show real reference images, while Rows 2 and 4 show the corresponding generated results. The viewing direction is controlled by the KML heading, and different columns correspond to different camera orientations.
Fig. 7 : Qualitative results of geo-semantic understanding and text-query geo-localization. (a) Given an image and its corresponding bounding boxes, UniGeo generates region-level grounded descriptions. (b) Each example presents the original query, the refined query, and the Top-10 retrieval results produced by the baseline and our method. Green boxes indicate correct candidates, and red boxes indicate incorrect candidates. (Best viewed when zooming in.)