Hierarchical visual localization pipeline

This work was done during my master's thesis under Prof. Leal-Taixe and Qunjie Zhou.

Visual localization pipeline using following steps

1. Find similar database images (neighbors) by using global descriptors
2. Extract local descriptors from neighboring database and query image
3. Match local descriptors
4. Calculate 6-DoF pose using RANSAC scheme

Overall concept

Current performance

Results

Evaluation via online evaluation system with benchmark results available.

GeM / Superpoint	Day	Night
High precision	71.0	31.6
Medium precision	79.5	46.9
Coarse precision	90.0	65.3

python evaluate.py --ratio_thresh 0.8 --reproj_error 14.0 --n_neighbors 20 --global_method Cirtorch --local_method Superpoint

GeM / SIFT	Day	Night
High precision	76.3	19.4
Medium precision	83.7	28.6
Coarse precision	87.7	36.7

Command to reproduce result:
python evaluate.py --ratio_thresh 0.75 --n_neighbors 20 --global_method Cirtorch

To use artificial night images mentioned in thesis you can download them here

Speed

Evaluated using following hardware:

• Intel(R) Xeon(R) CPU E5520 @ 2.27GHz
• GeForce GTX TITAN X

	Colmap	Superpoint
Setup time	50 seconds	45 seconds
Mean time / img	<1 seconds	3 seconds
Median time / img	<1 seconds	3 seconds
Max time / img	2 seconds	14 seconds

Get started

Prerequisites:

• Install conda
• Download AachenDayNight dataset
• (Optional) Install Colmap

Example for start on Linux

git clone https://github.com/a1302z/hierarchical_visual_localisation.git
cd hierarchical_visual_localisation
conda env create -f requirements.yml
mkdir data
cd data
wget https://syncandshare.lrz.de/dl/fiQXCXZ9ibmrm7rwUJzAvNL4
cd ..
mv <path to AachenDayNight dataset> data/

Credits

The concept of hierarchical localisation was introduced in this paper.

We used code from the following repositories.

• HF-Net
• Pytorch NetVLAD
• CNN Image Retrieval (GeM)
• Superpoint
• D2-Net
• PyTorch geometric

If we missed any credits please let us know.