UAV-Self-Positioning-23M-ZCN

Task: Image feature extraction / UAV self-positioning in low-altitude urban environments
Base model: timm/vit_small_patch16_224.augreg_in1k
Backbone: ViT-S
Library: timm, PyTorch
Dataset: Dmmm997/DenseUAV (DenseUAV)
Training Performance: Weights & Biases
License: Apache-2.0
Internship organization: Institute of Mathematical and Computational Sciences (IMACS), Ho Chi Minh City University of Technology, Vietnam
Supervisor: M.Sc. NGUYEN VAN GIA THINH

---

Model description

UAV-Self-Positioning-23M-ZCN is an image feature extraction model finetuned from the backbone timm/vit_small_patch16_224.augreg_in1k for vision-based UAV self-positioning in low-altitude urban environments.

This model is trained from the official DenseUAV baseline with a ViT-S backbone, using the code and training pipeline from the repository Dmmm1997/DenseUAV and the DenseUAV dataset (Dmmm997/DenseUAV). All credits for the dataset, baseline architecture, and evaluation protocol belong to the authors of DenseUAV. In this work I:

• configure baseline/opts.yaml for this experiment,
• train the model using the original baseline training scripts,
• and export the checkpoint as a model on Hugging Face.

The training performance logs are reported using Weights & Biases; check them at DenseUAV Non-GPS Training Performance

If you use this model in research or applications, please cite the DenseUAV paper and repository (see Citation below).

---

Origin and citation (important)

This model is fully based on:

• The original DenseUAV repository: Dmmm1997/DenseUAV
• The paper: Vision-Based UAV Self-Positioning in Low-Altitude Urban Environments (IEEE Transactions on Image Processing, 2024), arxiv.org/abs/2201.09201

If you use this model, please cite:

@misc{dai2023visionbaseduavselfpositioninglowaltitude,
      title={Vision-Based UAV Self-Positioning in Low-Altitude Urban Environments}, 
      author={Ming Dai and Enhui Zheng and Zhenhua Feng and Jiedong Zhuang and Wankou Yang},
      year={2023},
      eprint={2201.09201},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2201.09201}, 
}

And the repository:

Dmmm1997. DenseUAV: Vision-Based UAV Self-Positioning in Low-Altitude Urban Environments. GitHub, 2023. https://github.com/Dmmm1997/DenseUAV

I am not the author of DenseUAV; I only retrain and publish a checkpoint under the same Apache-2.0 license.

---

Quick usage

The model is published as an image feature extractor on Hugging Face, similar to the models under the Image Feature Extraction task on huggingface.co/models.

Example usage with timm (PyTorch):

• The uploaded checkpoint was saved from a wrapper model, so ViT weights are stored under the prefix backbone.backbone..
• The snippet below strips that prefix and loads the backbone successfully (prints torch.Size([1, 384])).

import os

import torch
from PIL import Image
from timm import create_model
from torchvision import transforms

device = "cuda" if torch.cuda.is_available() else "cpu"

# 1. Create ViT backbone
model = create_model(
    "vit_small_patch16_224",
    pretrained=False,
    num_classes=0,  # feature extractor
).to(device)

# 2. Download checkpoint from Hugging Face
state_dict = torch.hub.load_state_dict_from_url(
    "https://huggingface.co/Bancie/UAV-Self-Positioning-23M-ZCN/resolve/main/UAV_SelfPositioning_23M_ZCN.pth",
    map_location=device,
)

# 3. Strip wrapper prefix and load ViT weights
vit_prefix = "backbone.backbone."
vit_state_dict = {k[len(vit_prefix):]: v for k, v in state_dict.items() if k.startswith(vit_prefix)}
model.load_state_dict(vit_state_dict, strict=False)
model.eval()

# 4. Preprocess a local UAV-view image (224x224)
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)),
])

image_path = "/path/to/your/drone_image.png"
assert os.path.exists(image_path), f"Image not found: {image_path}"

image = Image.open(image_path).convert("RGB")
x = transform(image).unsqueeze(0).to(device)

# 5. Extract feature vector
with torch.no_grad():
    feat = model(x)  # [1, D]
    feat = torch.nn.functional.normalize(feat, dim=-1)

print(feat.shape)  # torch.Size([1, 384])

In the DenseUAV setting, these features are used to:

• build embeddings for UAV-view and satellite-view images,
• compute distances (e.g., cosine or Euclidean),
• and evaluate Recall and SDM as in the original DenseUAV code.

---

Training data

• Dataset: DenseUAV (Dmmm997/DenseUAV)
• Data type: UAV-view and satellite-view images captured in low-altitude urban environments
• Split: train / query / gallery as described in the original DenseUAV README
• Preprocessing: follows the baseline pipeline from the DenseUAV repository (resize, augmentation, normalization)

To fully reproduce training and evaluation, please clone the original repository and follow its instructions.

---

Training procedure

• Code: directly based on the baseline implementation under baseline/ in Dmmm1997/DenseUAV.
• Backbone: vit_small_patch16_224.augreg_in1k from timm.
• Config: a customized baseline/opts.yaml for this experiment.
• Scripts: trained using the original baseline scripts (for example train.py / train_test_local.sh) as suggested by the DenseUAV authors.

The architecture and loss functions follow the baseline; only minor hyperparameter choices and random seed are adapted to my compute resources.

---

Intended uses and limitations

Intended uses

• Research on UAV self-positioning and cross-view geo-localization.
• Backbone / feature extractor for new methods (e.g., variants or extensions built on top of DenseUAV).
• Quick experiments on DenseUAV without training from scratch.

Not intended for

• Out-of-domain UAV positioning (very different cities, weather conditions, altitudes, or sensors).
• Safety-critical applications without proper validation, calibration, and monitoring.

Limitations

• Trained on a specific dataset; may not generalize to all UAV scenarios.
• Processes single images; temporal information from video sequences is not explicitly modeled.

---

Evaluation

The evaluation protocol follows the original DenseUAV repository (Recall@K, SDM, etc.).
To re-evaluate this checkpoint on DenseUAV:

1. Clone the DenseUAV repository and prepare the dataset as described there.
2. Place this checkpoint under checkpoints/<name>/.
3. Run:

python test.py --name <name> --test_dir <dataset_root>/test
python evaluate_gpu.py
python evaluateDistance.py --root_dir <dataset_root>

Replace <name> with your checkpoint directory name and <dataset_root> with the path to the DenseUAV dataset. For details, please refer to the original DenseUAV README.

---

License

• Model checkpoint: released under Apache-2.0, the same license as the original DenseUAV project.
• Original code and dataset: owned by the DenseUAV authors – see the LICENSE file in Dmmm1997/DenseUAV.

By using this model, you agree to comply with the Apache-2.0 license of both DenseUAV and Hugging Face.