Awesome

Hierarchical Scene Coordinate Classification and Regression for Visual Localization

This is the PyTorch implementation of our paper, a hierarchical scene coordinate prediction approach for one-shot RGB camera relocalization:

Hierarchical Scene Coordinate Classification and Regression for Visual Localization, CVPR 2020
Xiaotian Li, Shuzhe Wang, Yi Zhao, Jakob Verbeek, Juho Kannala

Setup

Python3 and the following packages are required:

cython
numpy
pytorch
opencv
tqdm
imgaug

It is recommended to use a conda environment:

Install anaconda or miniconda.
Create the environment: conda env create -f environment.yml.
Activate the environment: conda activate hscnet.

To run the evaluation script, you will need to build the cython module:

cd ./pnpransac
python setup.py build_ext --inplace

Data

We currently support 7-Scenes, 12-Scenes, Cambridge Landmarks, and the three combined scenes which have been used in the paper. We will upload the code for the Aachen Day-Night dataset experiments.

You will need to download the datasets from the websites, and we provide a data package which contains other necessary files for reproducing our results. Note that for the Cambridge Landmarks dataset, you will also need to rename the files according to the train/test.txt files and put them in the train/test folders. And the depth maps we used for this dataset are from DSAC++. The provided label maps are obtained by running k-means hierarchically on the 3D points.

Evaluation

The trained models for the main experiments in the paper can be downloaded here.

To evaluate on a scene from a dataset:

python eval.py \
        --model [hscnet|scrnet] \
        --dataset [7S|12S|Cambridge|i7S|i12S|i19S] \
        --scene scene_name \
        --checkpoint /path/to/saved/model/ \
        --data_path /path/to/data/

Training

You can train the hierarchical scene coordinate network or the baseline regression network by running the following command:

python train.py \
        --model [hscnet|scrnet] \
        --dataset [7S|12S|Cambridge|i7S|i12S|i19S] \
        --scene scene_name \ # not required for the combined scenes
        --n_iter number_of_training_iterations \
        --data_path /path/to/data/

License

Acknowledgements

The PnP-RANSAC pose solver builds on DSAC++. The sensor calibration file and the normalization translation files for the 7-Scenes dataset are from DSAC. The rendered depth images for the Cambridge Landmarks dataset are from DSAC++.

Citation

Please consider citing our paper if you find this code useful for your research:

@inproceedings{li2020hscnet,
    title = {Hierarchical Scene Coordinate Classification and Regression for Visual Localization},
    author = {Li, Xiaotian and Wang, Shuzhe and Zhao, Yi and Verbeek, Jakob and Kannala, Juho},
    booktitle = {CVPR},
    year = {2020}
}