Domain-Invariant Feature Learning with Feature Consistency Loss (DIFL-FCL)

This is our Pytorch implementation for DIFL-FCL. Code was written by Hanjiang Hu based on ComboGAN

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

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}
}

Prerequisites

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

Getting Started

Installation

Install requisite Python libraries.

pip install torch
pip install torchvision
pip install visdom
pip install dominate
pip install scipy
pip install -U scikit-learn

Clone this repo:

git clone https://github.com/HanjiangHu/DIFL-FCL.git

Training

The dataset we use in this paper is from CMU-Seasons HERE. And it is then divided into CMU_urban, CMU_suburban and CMU_park area folders under ./datasets with different slices. In each area folder, all the images, including both database and query images, are categorized into 12 domains according to 12 environmental conditions, and placed in subfolders from train00 to train11 for training. And in order to retrieve images slice by slice while testing, the images of environmental condition X in slice Y are placed in sY/testXX as well. The groundtruth poses pose_new_sY.txt for database images are already in each slice folder.

Note that the number order of trainXX and testXX folders is strictly consistent with the environmental condition order introduced in README_CMU-Seasons.md, e.g. train00 and test00 correspond to the first environmental condition, i.e. the reference condition.

Some of our pretrained models for the CMU-Seasons dataset are found HERE. Please uncompress it under the root path. The pretrained models in cmu_urban, cmu_suburban and cmu_park folders under ./checkpoints are trained with L2-FCL only using images of each area. The urban models at 300 epoch for transfer learning are also given. It is also encouraged to use urban models to test suburban and park images in order to test the generation ability.

Examples:

Train a model:

python train.py --name cmu_urban --dataroot ./datasets/CMU_urban/ --n_domains 12 --niter 300 --niter_decay 300 --display_id 2

Checkpoints will be saved by default to ./checkpoints/cmu_urban/

Fine-tuning/Resume training:

python train.py --continue_train --which_epoch 300 --name cmu_urban --dataroot ./datasets/CMU_urban/ --n_domains 12 --niter 300 --niter_decay 300  --display_id 2

Test the model:

python test.py --phase test --name cmu_urban --dataroot ./datasets/CMU_urban/ --n_domains 12 --which_epoch 600 --serial_test --gpu_ids 0  --which_slice 2 --test_using_cos

The test result will be saved to txt file here: ./results/cmu_urban_600_s2_cos.txt.

Training/Testing 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; set --gpu_ids 0,1,2 for multi-GPU mode.
Visualization: during training the current image translation results and loss plots can be viewed by setting --display_id > 0. The results and loss plot will appear on a local graphics web server launched by visdom. To do this, you should have visdom installed and a server running by the command python -m visdom.server. The default server URL is http://localhost:8097. display_id corresponds to the window ID that is displayed on the visdom server.
Preprocessing: for training the default option resize_and_crop resizes the image such that the largest side becomes opt.loadSize and then does a random crop of size (opt.fineSize, opt.fineSize). While testing, images are only resized and not cropped. If --resize_to_crop_size option is specified, the image will be resiezed to size (opt.fineSize, opt.fineSize).
Feature loss metric: set loss weight --lambda_FCL to 0 to train without feature loss, otherwise specify --train_using_cos to use cosine distance as feature loss instead of L2, and --test_using_cos to use cosine similarity for image retrieval instead of L2.
About PCA: specify --test_after_pca to use PCA technique before retrieval, and the default dimension --PCA_dimension is 100. Note that since PCA is implemented through sklearn module, it takes more time in our demo due to lack of speedup with GPU.

HanjiangHu / DIFL-FCL