Dual-Resolution Correspondence Network In Python

Dual-Resolution Correspondence Network

Dual-Resolution Correspondence Network, NeurIPS 2020

Dependency

All dependencies are included in asset/dualrcnet.yml. You need to install conda
first, and then run

conda env create --file asset/dualrcnet.yml

To activate the environment, run

conda activate dualrcnet

Preparing data

We train our model on MegaDepth dataset. To prepare for the data, you need
to download the MegaDepth SfM models from the MegaDepth website and
download training_pairs.txt and validation_pairs.txt from this link.
Then place both training_pairs.txt and validation_pairs.txt files under the downloaded directory MegaDepth_v1_SfM.

Training

After downloading the training data, run

python train.py --training_file path/to/training_pairs.txt --validation_file path/to/validation_pairs.txt --image_path path/to/MegaDepth_v1_SfM

Pre-trained model

We also provide our pre-trained model. You can download dualrc-net.pth.tar from this link
and place it under the directory trained_models.

Evaluation on HPatches

The dataset can be downloaded from HPatches repo. You need to download
HPatches full sequences.
After downloading the dataset, then:

Browse to HPatches/
Run python eval_hpatches.py --checkpoint path/to/model --root path/to/parent/directory/of/hpatches_sequences. This will
generate a text file which stores the result in current directory.
Open draw_graph.py. Change relevent path accordingly and run the script to draw the result.

We provide results of DualRC-Net alongside with results of other methods in directory cache-top.

Evaluation on InLoc

In order to run the InLoc evaluation, you first need to clone the InLoc demo repo, and download and compile all the required depedencies. Then:

Browse to inloc/.
Run python eval_inloc_extract.py adjusting the checkpoint and experiment name.
This will generate a series of matches files in the inloc/matches/ directory that then need to be fed to the InLoc evaluation Matlab code.
Modify the inloc/eval_inloc_compute_poses.m file provided to indicate the path of the InLoc demo repo, and the name of the experiment (the particular directory name inside inloc/matches/), and run it using Matlab.
Use the inloc/eval_inloc_generate_plot.m file to plot the results from shortlist file generated in the previous stage: /your_path_to/InLoc_demo_old/experiment_name/shortlist_densePV.mat. Precomputed shortlist files are provided in inloc/shortlist.

Evaluation on Aachen Day-Night

In order to run the Aachen Day-Night evaluation, you first need to clone the Visualization benchmark repo, and download and compile all the required depedencies (note that you'll need to compile Colmap if you have not done so yet). Then:

Browse to aachen_day_and_night/.
Run python eval_aachen_extract.py adjusting the checkpoint and experiment name.
Copy the eval_aachen_reconstruct.py file to visuallocalizationbenchmark/local_feature_evaluation and run it in the following way:

python eval_aachen_reconstruct.py 
	--dataset_path /path_to_aachen/aachen 
	--colmap_path /local/colmap/build/src/exe
	--method_name experiment_name

Upload the file /path_to_aachen/aachen/Aachen_eval_[experiment_name].txt to https://www.visuallocalization.net/ to get the results on this benchmark.

BibTex

If you use this code, please cite our paper

@inproceedings{li20dualrc,
 author		= {Xinghui Li and Kai Han and Shuda Li and Victor Prisacariu},
 title   	= {Dual-Resolution Correspondence Networks},
 booktitle 	= {Conference on Neural Information Processing Systems (NeurIPS)},
 year    	= {2020},
}