Research map / Detection and segmentation / Open-set and few-shot detection
Locate Anything on Earth: Advancing Open-Vocabulary Object Detection for Remote Sensing Community
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Reformulates open-vocabulary detection for remote sensing as Locate Anything on Earth, builds the LAE-1M dataset via a semi-automated label engine (SAM + InternVL), and proposes LAE-DINO with dynamic vocabulary construction and visual-guided text prompt learning.
Problem
Open-vocabulary object detectors trained on natural images generalize poorly to remote sensing imagery due to domain gap, and remote sensing lacks large, diverse object-level labeled data with a large vocabulary.
Approach
A label engine that unifies sliced labeled RS detection datasets and auto-annotates unlabeled RS imagery with SAM plus an LVLM, combined with a GroundingDINO-style detector that samples a fixed-size vocabulary per batch (DVC) and adds a visual-derived scene feature to text prompts (VisGT).
Key results
- LAE-DINO reaches 85.5 AP50 on DIOR and 46.8 mAP on DOTAv2.0 in open-set settings, versus 83.6 and 46.0 for GroundingDINO with DVC.
- After closed-set fine-tuning on DOTAv2.0, LAE-DINO-FT with LAE-1M pre-training reaches 57.9 mAP versus 54.0 for GroundingDINO-FT (O365,GoldG,Cap4M).
- Zero-shot GroundingDINO and GLIP trained on natural data score near 0 on remote sensing benchmarks, e.g. 0.3 AP50 on DIOR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add VisGT during LAE-1M pre-training (ablation) | LAE-DINO · DIOR open-set | AP50: 83.6 -> 85.5 | Helped |
| Add VisGT during fine-tuning (ablation) | LAE-DINO · DIOR fine-tuning | AP50: 89.9 -> 92.0 | Helped |
| Add LAE-1M pre-training to VisGT fine-tuning (ablation) | LAE-DINO · DIOR fine-tuning | AP50: 92.0 -> 92.2 | Helped |
| Add LAE-COD to LAE-FOD pre-training (ablation) | LAE-DINO · DIOR open-set | AP50: 84.1 -> 85.5(+1.4) | Helped |
| Add LAE-COD to LAE-FOD pre-training (ablation) | LAE-DINO · DOTAv2.0 open-set | mAP: 44.5 -> 46.8(+2.3) | Helped |
| Add LAE-COD to LAE-FOD pre-training (ablation) | LAE-DINO · LAE-80C open-set | mAP: 19.1 -> 20.2(+1.1) | Helped |
| Pre-train LAE-DINO on LAE-1M instead of natural-scene data | LAE-DINO · DIOR open-set | AP50: 83.6 vs 0.3 | Helped |
| Use LAE-DINO instead of GroundingDINO with DVC | Open-vocabulary remote-sensing detection · DOTAv2.0 open-set | mAP: +0.8% | Helped |
| Use LAE-DINO instead of GroundingDINO with DVC | Open-vocabulary remote-sensing detection · LAE-80C open-set | mAP: +2.5% | Helped |
| Pre-train LAE-DINO on LAE-1M rather than O365, GoldG, Cap4M | LAE-DINO-FT · DOTAv2.0 fine-tuning | mAP: 55.5 -> 57.9 | Helped |
| Reduce fine-tuning data from DIOR-full to DIOR-½ | LAE-DINO · DIOR test | AP50: 92.2 -> 89.1 | Hurt |
Surprising findings
- GroundingDINO and GLIP pretrained on natural-scene data score near zero on DIOR and DOTAv2.0 in open-set evaluation.
Limitations
- Annotation quality of LAE-COD is only assessed with a small user survey of ten samples, and box quality is rated lower than category quality.
- Closed-set comparison generic detectors are cited from original papers because their code is not open source.
- Novel class evaluation is limited because some rare classes lack quantitative ground truth.
Experimental setup
Datasets: LAE-1M DOTA DIOR FAIR1M NWPU VHR-10 RSOD xView HRSC2016 Condensing-Tower AID NWPU-RESISC45 SLM
Benchmarks: DIOR DOTAv2.0 LAE-80C HRRSD
Metrics: AP50 mAP mAP_base mAP_novel
Baselines: GLIP GroundingDINO GLIP with DVC GroundingDINO with DVC GASSL CACO TOV Scale-MAE SatLas RingMo SkySense MTP
Builds on: GroundingDINO DINO MQ-Det SAM
Closest papers by meaning
- InstructSAM: A Training-free Framework for Instruction-Oriented Remote Sensing Object Recognition (NeurIPS 2025 · similarity 0.92)
- SkySense-O: Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling (CVPR 2025 · similarity 0.89)
- SkyScript: A Large and Semantically Diverse Vision-Language Dataset for Remote Sensing (AAAI 2024 · similarity 0.88)
- VRSBench: A Versatile Vision-Language Benchmark Dataset for Remote Sensing Image Understanding (NeurIPS 2024 · similarity 0.87)
- Towards Open-Vocabulary Remote Sensing Image Semantic Segmentation (AAAI 2025 · similarity 0.87)
- TEOChat: A Large Vision-Language Assistant for Temporal Earth Observation Data (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.