czslxk / dasgil

DASGIL: Domain Adaptation for Semantic and Geometric-aware Image-based Localization

This is our Pytorch implementation for DASGIL (arxiv) by Hanjiang Hu, Zhijian Qiao and Ming Cheng.

Prerequisites

• Linux or macOS
• Python 3
• CPU or NVIDIA GPU + CUDA CuDNN

Getting Started

Installation

• Install requisite Python libraries.

pip install -r requirements.txt

• Clone this repo:

git clone https://github.com/HanjiangHu/DASGIL.git

Training

KITTI and Virtual KITTI 2 dataset are used to train the model, while Extended CMU-Seasons dataset is used to test. The datasets involved in this paper are well organized HERE. Please uncompress it under the root path. Our pretrained models are found HERE. Please uncompress it under the root path.

• Training on KITTI and Virtual KITTI Dataset:

python train.py --name DASGIL

• Fine-tune the pretrained model:

python train.py --name DASGIL --continue_train --which_epoch 200

Testing

• Testing on the Extended CMU-Seasons Dataset:

python test.py --name DASGIL --which_epoch 200

Results

The test results will be saved to ./output. The txt results should be merged into a single txt file and submitted to the official benchmark website.

Our DASGIL results on Extended CMU-Seasons Dataset could be found on the benchmark website.

Other Details

• See ./options/train_options.py for training-specific flags, ./options/test_options.py for test-specific flags, and ./options/base_options.py for all common flags.
• CPU/GPU (default --gpu_ids 0): set--gpu_ids -1 to use CPU mode (NOT recommended); set --gpu_ids 0,1,2 for multi-GPU mode.

If you use this code in your own work, please cite:

H. Hu, M. Cheng, Z. Liu and H. Wang ”DASGIL: Domain Adaptation for Semantic and Geometric-aware Image-based Localization”,

@misc{hu2020dasgil,
      title={DASGIL: Domain Adaptation for Semantic and Geometric-aware Image-based Localization}, 
      author={Hanjiang Hu and Ming Cheng and Zhe Liu and Hesheng Wang},
      year={2020},
      eprint={2010.00573},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}