Framework

🎯 SCC-Loc is a zero-shot, cross-modal thermal geo-localization framework for UAVs operating in GNSS-denied environments. Matching onboard thermal images with visible-light satellite maps is challenging due to the massive modality gap. SCC-Loc tackles this by providing highly accurate absolute pose estimation without needing domain-specific training .

✨ Key Highlights:

• Zero-Shot Capability: Powered by a shared DINOv2 backbone, requiring no specific retraining for new environments.
• Robust Matching: Overcomes dense structural outliers and visual decoys.
• Unified Framework: Integrates Semantic-Guided Viewport Alignment, Spatial-Adaptive Filtering, and Consensus-Driven Position Selection.

Dataset

🀄 We constructed the Thermal-UAV dataset, comprising 11,890 thermal UAV images captured using a DJI Matrice 4T drone around Changsha. These images feature nadir views, multi-temporal and multi-scenario data, and continuous flight sequences. We split the dataset into three sets: 8,115 for training, 1,425 for validation, and 2,350 for testing. Additionally, we collected corresponding Google Maps satellite imagery at 0.26 m/px and a Digital Surface Model (DSM) at 5.29 m/px. The directory structure of the Thermal-UAV dataset is as follows:

Data/
├── metadata/
│   ├── train_Thermal.json
│   ├── valid_Thermal.json
│   └── test_Thermal.json
├── Reference_map/
│   └── changsha/
│       ├── ref.tif
│       └── dsm.tif
└── Thermal-UAV/
  ├── train/
  │   └── changsha/
  │       └── <NUDT>/
  │           └── Thermal/
  │               ├── xxx1.JPG
  │               ├── xxx2.JPG
  |	          └──  Thermal_info.csv
  │       └── ...
  ├── valid/
  └── test/

The Thermal-UAV dataset and checkpoint are provided in Baidu Netdisk and Hugging Face. You can use the process.ipynb to get the metadate of Thermal-UAV JSON format, as follows:

{
  "name": "./Data/Thermal-UAV/train/changsha/city_300_ortho_night/Thermal/xxx_T.JPG",
  "lat": 28.2436611, # latitude
  "lon": 112.9985009, # longitude
  "abs_height": 341.354, # absolute height
  "rel_height": 323.967, # UAV-to-ground distance measured by laser rangefinder (direct use this character due to the vertical angle of view)
  "pitch": -90.0, # pitch
  "yaw": 88.0, # yaw
  "roll": 180.0, # roll
  "cam_size": 9.83, # diagonal physical size of the sensor
  "focal_len": 12.0, # focal length (mm)
  "width": 1280.0, # image width
  "height": 1024.0 # image height
}

It is noted that we need to fill in the corresponding changsha geographic infomation in Regions_params/, as follows:

changsha_UTM_SYSTEM: 49N
changsha_SAMPLE_INTERVAL: 10
changsha_HIGH_REF_PATH: ./Data/Reference_map/changsha/ref.tif
changsha_HIGH_DSM_PATH: ./Data/Reference_map/changsha/dsm.tif
changsha_HIGH_REF_initialX: 694811.4577 # UTM coordinate
changsha_HIGH_REF_initialY: 3130673.6615 # UTM coordinate
changsha_HIGH_REF_resolution: 0.3 # Align the resolution through upsampling on QGIS
changsha_HIGH_DSM_resolution: 0.3 # Align the resolution through upsampling on QGIS
changsha_HIGH_REF_COORDINATE: # If the satellite image and DSM are aligned. Their offset is 0
  - 0.0
  - 0.0
changsha_HIGH_DSM_COORDINATE: 
  - 0.0
  - 0.0

Checkpoints

1. Retrieval Model

• CAMP

Retrieval_Models/
└── CAMP/
  └── weights/
    └── weights_0.9446_for_U1652.pth

• DINOv3 and DINOv2 are offered through torch.hub. They will automatically download if your internet connection is available .

2. Matching Model

• RoMa

Matching_Models/
└── RoMa/
    └── ckpt/
        ├── roma_outdoor.pth
        └── dinov2_vitl14_pretrain.pth

• MINIMA

Matching_Models/
└── MINIMA/
    └── weights/
        ├── minima_eloftr.pth
        ├── minima_roma.pth
        ├── minima_loftr.ckpt
        ├── minima_lightglue.pth
        └── minima_xoftr.ckpt

• RoMaV2 automatically download

Method

In config.yaml， choose the MINIMA_Roma_DINOv2 as retrieval method and MINIMA_Roma as matching method. Our proposed SCC-Loc framework composes this combination. It is noted that when use MINIMA_Roma_DINOv2, it must use MINIMA_Roma, while the reverse is not true. Other combinations can be the comparision baseline.

Then, enjoy the fun of operation through:

python Baseline.py

If you want to see the visualization, please set the SHOW_RETRIEVAL_RESULT=True in config.yaml. It will show retrieval, matching, final localization resules etc.

🚀 Furthermore , process.ipynb provides useful scripts for constructing custom datasets and reproducing experimental results. Given the current strict regulations and flight restrictions on UAVs , we warmly invite researchers dedicated to thermal geo-localization to contribute to this open-source dataset, expanding its coverage to encompass diverse countries, regions, environments, and viewpoints .

Citation

If you find this code or our Thermal-UAV dataset useful for your research, please consider citing our paper:

@article{zhang2026scc,
  title={SCC-Loc: A Unified Semantic Cascade Consensus Framework for UAV Thermal Geo-Localization},
  author={Zhang, Xiaoran and Liu, Yu and Liang, Jinyu and Li, Kangqiushi and Huang, Zhiwei and Xiao, Huaxin},
  journal={arXiv preprint arXiv:2604.03120},
  year={2026}
}

License

We use the Apache License 2.0. See detailed information in LICENSE file.

Acknowledgements

We are grateful for the publicly available resources and open-source libraries that have been instrumental in this work.

• A Cross-View Geo-Localization Method using Contrastive Attributes Mining and Position-aware Partitioning
• DINOv2: Learning Robust Visual Features without Supervision
• DINOv3
• RoMa: Robust Dense Feature Matching
• RoMa v2: Harder Better Faster Denser Feature Matching
• MINIMA: Modality Invariant Image Matching

We are especially grateful to Yibin Ye et al. for their seminal benchmark Exploring the best way for UAV visual localization under Low-altitude Multi-view Observation Condition: a Benchmark

Thank you for your open-source spirit, which has significantly accelerated progress in the UAV visual geo-localization community. Salute to you!