topo-epfl / crossloc

[CVPR'22] CrossLoc localization: a cross-modal visual representation learning method for absolute localization

CrossLoc localization: a cross-modal visual representation learning method for absolute localization

This repository contains the official implementation of the CrossLoc localization algorithm.

Please make sure you have access to the CrossLoc Benchmark Raw Datasets and have set it up properly before proceeding.

Also check out other useful repos regarding the datasets:

• CrossLoc-Benchmark-Datasets: CrossLoc benchmark datasets setup tutorial.
• TOPO-DataGen: an open and scalable aerial synthetic data generation workflow.
• DDLoc: our adapation of ARC method for sim2real coordinate regression

Pretrained network weights, the testing set images and the full datasets could be found as follows:

• oneDrive
• Google Drive
• Dryad (Full CrossLoc Benchmark Datasets only)

Happy coding! :)

The CrossLoc localization algorithm is officially presented in the paper accepted to CVPR 2022
CrossLoc: Scalable Aerial Localization Assisted by Multimodal Synthetic Data
Qi Yan, Jianhao Zheng, Simon Reding, Shanci Li, Iordan Doytchinov
École Polytechnique Fédérale de Lausanne (EPFL)
Links: website | arXiv | code repos | datasets

Get started

Install dependencies

• If conda environment is available:

conda env create -f setup/environment.yml
conda activate crossloc

cd dsacstar && python3 setup_super.py --conda
# sanity check for DSAC* plugin
python3 -c "import torch; import dsacstar; print('DSAC* installation is fine')"

Note: import torch must be used before import dsacstar in the python script. conda environment is preferred as it handles the low-level opencv dependencies quite easily.

• Otherwise, if conda environment is not readily available:

python3 -m venv venvcrossloc
source venvcrossloc/bin/activate
pip3 install pip -U && pip3 install -r setup/requirements.txt

wget -O opencv-3.4.2.zip https://github.com/opencv/opencv/archive/refs/tags/3.4.2.zip
unzip -q opencv-3.4.2.zip && rm opencv-3.4.2.zip
mkdir -p opencv-build && cd opencv-build
cmake -DCMAKE_INSTALL_PREFIX=install ../opencv-3.4.2
cmake --build . -j12 --target install 
rm -rf ../opencv-3.4.2

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$(pwd)/install/lib:$(pwd)/install/lib64
cd ../dsacstar && python3 setup_super.py --cv_path ../opencv-build/install
# sanity check for DSAC* plugin
python3 -c "import torch; import dsacstar; print('DSAC* installation is fine')"

Note:

• Run export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$(pwd)/install/lib:$(pwd)/install/lib64 or equivalent commands in shell to add opencv lib directory EACH time before import dsacstar .

• Due to an issue in pytorch 1.9.0 stable release, some redundant warning messages may be popped out in the terminal (Warning: Leaking Caffe2 thread-pool after fork). Update to nightly pytorch may solve the issue:

  pip3 uninstall torch torchvision torchaudio
  pip3 install pip -U
  pip3 install --pre torch torchvision torchaudio -f https://download.pytorch.org/whl/nightly/cu102/torch_nightly.html

Setup datasets

See CrossLoc Benchmark Datasets Setup for details on dataset setup tutorials.

Training and testing

We propose to 1) pretrain the networks using task-agnostic synthetic data only and then 2) fine-tune the models on specific downstream tasks.

To be concise, we only show the steps for naturescape datasets in the main README. Please refer to another note detailing the other training, validation and testing steps for urbanscape datasets and ablation studies.

Encoders Pretraining

• Training: task-agnostic LHS-sim synthetic data is used (at train_sim folder).

# specify checkpoint weight output path
export CKPT_DIR=$(pwd)/ckpt-weights

# pretrain encoders with LHS-sim data for naturescape
bash script_clean_training/encoder_pretrain.sh naturescape coord FULL 1.0 in_place 0.0 mle 0
bash script_clean_training/encoder_pretrain.sh naturescape depth FULL 1.0 in_place 0.0 mle 0
bash script_clean_training/encoder_pretrain.sh naturescape normal FULL 1.0 in_place 0.0 mle 0
bash script_clean_training/encoder_pretrain.sh naturescape semantics FULL 1.0 in_place 0.0 none 0

• Checkpoint selection: we evaluate the model performance on the validation set (at val_sim folder) and select the checkpoint models for later training tasks.

# specify checkpoint weight output path
export CKPT_DIR=$(pwd)/ckpt-weights/$TASK_DIR
# please specify $TASK_DIR for each task, e.g., naturescape-coord-sclean_training-unc-MLE-e100-lr0.0002-sim_only-sc1.00
# otherwise, the validation script may not load the network weight properly

# select model weight based on validation set performance for naturescape
bash script_clean_validation/validate_encoder_pretrain.sh naturescape coord FULL mle 0
bash script_clean_validation/validate_encoder_pretrain.sh naturescape depth FULL mle 0
bash script_clean_validation/validate_encoder_pretrain.sh naturescape normal FULL mle 0
bash script_clean_validation/validate_encoder_pretrain.sh naturescape semantics FULL none 0
# select the checkpoint from the generated path, see script_clean_validation/select_ckpt.py for details

Encoders Fine-tuning

• Training: to fine-tune the encoders with real-synthetic paired date. Note that the pretrained encoders' weights must be spcified in the script. Check the variable ENC_PRETRAINED in the encoder_finetune.sh script for detailed setup.

# specify checkpoint weight output path
export CKPT_DIR=$(pwd)/ckpt-weights

# finetune encoders with in-place sim-to-real pairs for naturescape [using 100% LHS-pretrained weights]
bash script_clean_training/encoder_finetun.e.sh naturescape coord FULL 1.0 in_place 1.0 mle 0
bash script_clean_training/encoder_finetune.sh naturescape depth FULL 1.0 in_place 1.0 mle 0
bash script_clean_training/encoder_finetune.sh naturescape normal FULL 1.0 in_place 1.0 mle 0
bash script_clean_training/encoder_finetune.sh naturescape semantics FULL 1.0 in_place 1.0 none 0

# finetune encoders with out-of-place sim-to-real pairs for naturescape [using 100% LHS-pretrained weights]
bash script_clean_training/encoder_finetune.sh naturescape coord FULL 1.0 out_of_place 1.0 mle 0
bash script_clean_training/encoder_finetune.sh naturescape depth FULL 1.0 out_of_place 1.0 mle 0
bash script_clean_training/encoder_finetune.sh naturescape normal FULL 1.0 out_of_place 1.0 mle 0
bash script_clean_training/encoder_finetune.sh naturescape semantics FULL 1.0 out_of_place 1.0 none 0

• Checkpoint selection: again, we evaluate the model performance on the validation set (now at val_drone_real folder) and select the checkpoint models for later training tasks.

# specify checkpoint weight output path
export CKPT_DIR=$(pwd)/ckpt-weights/$TASK_DIR
# please specify $TASK_DIR for each task, e.g., naturescape-coord-sclean_training_pt1.00-unc-MLE-e800-lr0.0001-pairs-ip-rc1.00-finetune
# otherwise, the validation script may not load the network weight properly

# select model weight based on validation set performance naturescape data
# please change the $TASK_DIR and repeat for in-place and out-of-place scenes
export MIN_CKPT_ITER=1000000
bash script_clean_validation/validate_encoder_finetune.sh naturescape coord FULL mle 0
bash script_clean_validation/validate_encoder_finetune.sh naturescape depth FULL mle 0
bash script_clean_validation/validate_encoder_finetune.sh naturescape normal FULL mle 0
bash script_clean_validation/validate_encoder_finetune.sh naturescape semantics FULL none 0
# select the checkpoint from the generated path, see script_clean_validation/select_ckpt.py for details

Decoders Fine-tuning

• Training: to reuse the multiple fine-tuned encoders and fine-tune the decoder with real-synthetic paired date. Note that the encoders' weights from the last step must be spcified in the script. Check the variable ENC_COORD, ENC_DEPTH, ENC_NORMAL and ENC_SEMANTICS in the decoder_finetune.sh script for detailed setup.

# specify checkpoint weight output path
export CKPT_DIR=$(pwd)/ckpt-weights

# finetune decoder with in-place sim-to-real pairs for naturescape 
# [using 100% LHS-pretrained + sim-to-real paired data fine-tuned encoders]
bash script_clean_training/decoder_finetune.sh naturescape coord FULL 1.0 in_place 1.0 0.0 in_place 1.0 mle 0
bash script_clean_training/decoder_finetune_plus_semantics.sh naturescape coord FULL 1.0 in_place 1.0 0.0 in_place 1.0 mle 0

# finetune decoder with out-of-place sim-to-real pairs for naturescape 
# [using 100% LHS-pretrained + sim-to-real paired data fine-tuned encoders]
bash script_clean_training/decoder_finetune.sh naturescape coord FULL 1.0 out_of_place 1.0 0.0 out_of_place 1.0 mle 0
bash script_clean_training/decoder_finetune_plus_semantics.sh naturescape coord FULL 1.0 out_of_place 1.0 0.0 out_of_place 1.0 mle 0

• Checkpoint selection: again, we evaluate the model performance on the validation set (now at val_drone_real folder) and select the checkpoint models for later training tasks.

# specify checkpoint weight output path
export CKPT_DIR=$(pwd)/ckpt-weights/$TASK_DIR
# please specify $TASK_DIR for each task, e.g., naturescape-coord-decoder_coord_free_depth_normal-senc-pt1.00-ip-ft1.00-unc-MLE-e1000-lr0.0001-pairwise-ip-rc1.00
# otherwise, the validation script may not load the network weight properly

# select model weight based on validation set performance naturescape data
# please change the $TASK_DIR and repeat for in-place and out-of-place scenes
export MIN_CKPT_ITER=1000000  # in-place
export MIN_CKPT_ITER=200000   # out-of-place
bash script_clean_validation/validate_decoder_finetune.sh naturescape coord FULL mle 0
# select the checkpoint from the generated path, see script_clean_validation/select_ckpt.py for details

Performance Testing

Using the model weight selected above, we now run the testing script to evaluate the model's final performance on testing set.

# specify the specific weight path, change this accordingly before running each line
export WEIGHT_PATH=YOUR_PATH

# naturescape, in-place scene
python3 test_single_task.py naturescape --task coord --uncertainty mle --section test_drone_real --network_in ${WEIGHT_PATH}
# naturescape, out-of-place scene
python3 test_single_task.py naturescape --task coord --uncertainty mle --section test_oop_drone_real --network_in ${WEIGHT_PATH}

Citation

If you find our code useful for your research, please cite the paper:

@article{yan2021crossloc,
  title={CrossLoc: Scalable Aerial Localization Assisted by Multimodal Synthetic Data},
  author={Yan, Qi and Zheng, Jianhao and Reding, Simon and Li, Shanci and Doytchinov, Iordan},
  journal={arXiv preprint arXiv:2112.09081},
  year={2021}
}

@misc{iordan2022crossloc, 
	title={CrossLoc Benchmark Datasets}, 
	author={Doytchinov, Iordan and Yan, Qi and Zheng, Jianhao and Reding, Simon and Li, Shanci}, 
	publisher={Dryad}, 
	doi={10.5061/DRYAD.MGQNK991C}, 
	url={http://datadryad.org/stash/dataset/doi:10.5061/dryad.mgqnk991c},
	year={2022}
}

crossloc's Issues

Question about the reproduced result

Hi,

Thanks for the code, I'm very interested in your outstanding work!!!
I have tried to retrain your network, and the result shows that the index values of in-place localization accuracy I reproduced over urbanscape dataset are better than those given in your paper. So I want to ask you what might cause this situation.

Looking forward to your reply.

---

Here is the result over urbanscape dataset from your paper:

Here is quantitative comparison between paper's result and my reproduced result:

self.split, scene = split.split('_') ValueError: not enough values to unpack (expected 2, got 1)