DISAM: Domain-invariant Similarity Activation Map Metric Learning for Retrieval-based Long-term Visual Localization
This is our Pytorch implementation for DISAM (paper) as an extended version of DIFL-FCL (paper,code) in IROS 2019. The journal paper has been published in IEEE/CAA Journal of Automatica Sinica.
- Linux or macOS
- Python 3
- CPU or NVIDIA GPU + CUDA CuDNN
- Install requisite Python libraries.
pip install torch
pip install torchvision
pip install visdom
pip install dominate- Clone this repo:
git clone https://github.com/HanjiangHu/DISAM.gitThe datasets used to train and test in this paper are well organized HERE. Please uncompress it under the root path.
Some of our pretrained models for the CMU-Seasons dataset are found HERE. Please make a new directory ./checkpoints under the root path and uncompress it under ./checkpoints. The pre-trained models for RobotCar Dataset are originated from models from CMU folders and have been already transferred according to the environment correspondence in Table Ⅰ of the paper. The pre-trained image translation models at epoch 300 could be moved to any other folder to fine-tune using Urban images of CMU-Seasons Dataset.
Training on Urban area of CMU-Seasons Dataset:
- Fine-tune a coarse model from epoch 300:
python train.py --name CMU_coarse --dataroot the/path/to/CMU_urban --n_domains 12 --niter XXX --niter_decay XXX --gpu_ids 0 --lambda_sam 0.0 --continue_train --which_epoch 300
- Fine-tune a fine model from epoch 300:
python train.py --name CMU_fine --dataroot the/path/to/CMU_urban --n_domains 12 --niter XXX --niter_decay XXX --gpu_ids 0 --train_using_cos --mean_cos --use_cos_latent_with_L2 --continue_train --which_epoch 300
Testing on CMU-Seasons Dataset:
- Test the coarse model:
python test.py --phase test --name CMU_coarse --dataroot the/path/to/CMU_AREA --n_domains 12 --which_epoch 1200 --serial_test --gpu_ids 0 --which_slice XX --test_using_cos --mean_cos
- Test the fine model:
python test.py --phase test --name CMU_fine --dataroot the/path/to/CMU_AREA --n_domains 12 --which_epoch 1200 --serial_test --gpu_ids 0 --which_slice XX --test_using_cos
- Test the coarse-to-fine pipeline:
python test.py --phase test --name CMU_coarse --dataroot the/path/to/CMU_AREA --n_domains 12 --which_epoch 1200 --use_two_stage --top_n 3 --name_finer CMU_fine --which_epoch_finer 1200 --serial_test --gpu_ids 0 --which_slice XX --test_using_cos --mean_cos
Testing on RobotCar Dataset:
- Build the feature database:
python save_database_feature.py --phase test --name robotcar_coarse --dataroot the/path/to/RobotCar_rear --n_domains 10 --which_epoch 1200 --serial_test --gpu_ids 0 --test_using_cos --mean_cos --use_two_stage --name_finer robotcar_fine --which_epoch_finer 1200
- Test the coarse model:
python test_robotcar.py --phase test --name robotcar_coarse --dataroot the/path/to/RobotCar_rear --n_domains 10 --which_epoch 1200 --serial_test --gpu_ids 0 --test_using_cos --test_condition XX --mean_cos
- Test the fine model:
python test_robotcar.py --phase test --name robotcar_fine --dataroot the/path/to/RobotCar_rear --n_domains 10 --which_epoch 1200 --serial_test --gpu_ids 0 --test_using_cos --only_for_finer --resize64 --test_condition XX --mean_cos
The test results will be saved to the root path. The txt results should be merged into a single txt file and submitted to the official benchmark website.
Our coarse-only, fine-only and coarse-to-fine results on CMU-Seasons Dataset could be found on the benchmark website. Our coarse-only and fine-only results on RobotCar Dataset could be found on the benchmark website as well.
- See
options/train_options.pyfor training-specific flags,options/test_options.pyfor test-specific flags on CMU-Seasons Dataset,options/RobotcarTestOptions.pyfor test-specific flags on RobotCar Dataset, andoptions/base_options.pyfor all common flags. - CPU/GPU (default
--gpu_ids 0): set--gpu_ids -1to use CPU mode; set--gpu_ids 0,1,2for multi-GPU mode. - More details could be found in DIFL-FCL
If you use this code in your own work, please cite:
H. Hu, H. Wang, Z. Liu, and W. Chen ”Domain-invariant Similarity Activation Map Metric Learning for Retrieval-based Long-term Visual Localization”,
@ARTICLE{hu2021domain,
author={H. {Hu} and H. {Wang} and Z. {Liu} and W. {Chen}},
journal={IEEE/CAA Journal of Automatica Sinica},
title={Domain-invariant similarity activation map metric learning for retrieval-based long-term visual localization},
year={2021},
volume={},
number={},
pages={1-16},
doi={10.1109/JAS.2021.1003907}}
H. Hu, H. Wang, Z. Liu, C. Yang, W. Chen, and L. Xie
”Retrieval-based Localization Based on Domain-invariant Feature Learning under Changing Environments”,
IROS 2019
@inproceedings{hu2019DIFLFCL,
author={H. {Hu} and H. {Wang} and Z. {Liu} and C. {Yang} and W. {Chen} and L. {Xie}},
booktitle={2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
title={Retrieval-based Localization Based on Domain-invariant Feature Learning under Changing Environments},
year={2019},
pages={3684-3689}
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent