Chen, Chao and Li, Chunyu and He, Mengfan and Wang, Jun and Xing, Fei and Meng, Ziyang
-
[✅] The ONNX model and its export code will be released.
-
The optimized ONNX export code enables the TensorRT model to run at approximately 44 Hz on the Jetson Orin NX, significantly surpassing the 14 Hz reported in the paper.
-
The model weights trained on a larger dataset will be released.
- mmcv
pip install -U openmim
mim install mmcv==1.7.2
- faiss
conda install -c pytorch faiss-cpu
You can download the training data from Baidu Netdisk or Tsinghua Cloud.
You can download the test data from here. VPAir, UAV-VisLoc (eval): Baidu Netdisk or Tsinghua cloud
If you need the raw data from UAV-VisLoc
UltraVPR's weights: Baidu Netdisk or Tsinghua cloud
If you want to retrain the model, please first load the pre-trained weights of the backbone network. Relevant code is in utils.py.
E2ResNet's weights: Baidu Netdisk or Tsinghua cloud
After preparing the dataset, please modify the path in dataloader/aerialvl.py.
python main.py --mode=train
If you want to use the model for enhanced training:
python main.py --mode=train --upscaling=True
If you want to test on VPAir, please run the following code:
python eval_VPAir.py
If you want to test on UAV-VisLoc, please run the following code:
python eval_UAV-VisLoc.py
This work draws inspiration from the following code as references. We extend our gratitude to these remarkable contributions:
When using code within this repository, please refer the following paper in your publications:
@ARTICLE{11091472,
author={Chen, Chao and Li, Chunyu and He, Mengfan and Wang, Jun and Xing, Fei and Meng, Ziyang},
journal={IEEE Robotics and Automation Letters},
title={UltraVPR: Unsupervised Lightweight Rotation- Invariant Aerial Visual Place Recognition},
year={2025},
volume={10},
number={9},
pages={9096-9103}}