Deep Retrieval and image Alignment Network

We introduce DRAN, a neural network able to perform all the task present in Visual Localization. DRAN solves Visual Localization in three stages, using the features from a single network:

• Image Retrieval: global descriptors are extracted by a GeM pooling stage. Map features are compared against the query descriptor with a dot product.
• Re-ranking and initial pose: we extract hypercolumns from the encoder features, applying the Sparse-to-Dense method. This step filters the candidates and obtains an initial pose estimation for the query image.
• Camera pose refinement: we use PixLoc to align the features extracted from our trained decoder, by means of feature-metric optimization.

Installation

DRAN is built with Python >=3.6 and PyTorch. To install it, run:

git clone https://github.com/jmorlana/DRAN
cd DRAN/
pip install -e .

Paths are defined in dran/settings.py. Folder structure is as follows:

.
├── datasets     # public datasets
└── outputs
    ├── training # checkpoints and training logs
    ├── hloc     # 3D models and retrieval for localization
    └── results  # outputs of the evaluation

Download trained models and evaluation data

Before starting, you have to download the trained models, which can be found here.

Place retrieval models (from the Google Drive link above) under dran/retrieval/checkpoints and localization models under outputs/training.

To download the evaluation data of the 3 datasets reported, run the following command:

python -m dran.download --select Aachen CMU RobotCar

Usage

Despite using the same underlying network weights, retrieval and localization are implemented separately just for simplicity. Retrieval can be repoduced for the different datasets. For example, for Aachen:

cd DRAN/dran
python -m retrieval.Aachen

Localization results, perfoming re-ranking and pose optimization can be reproduced by calling:

cd DRAN
python -m dran.run_Aachen

Training

Data preparation

The 3D point clouds, camera poses, and intrinsic parameters are preprocessed together to allow for fast data loading during training. These files are generated using the scripts dran/lib/preprocess_[cmu|megadepth].py. Data can be download via:

python -m dran.download --select CMU MegaDepth --training

This also downloads the training split of the CMU dataset. The undistorted MegaDepth data (images) can be downloaded from the D2-Net repository. For more details about training data, please refer to PixLoc.

Training experiment

A training experiment can be launched using the following command:

python -m dran.dran.train megadepth_test \
		--conf dran/lib/configs/dran_megadepth.yaml

It will initialize a frozen encoder with the retrieval weights stored in dran/retrieval/checkpoints, fine-tuning the localization head, the decoder and the pose optimizer.

Implementation

This repository is a modified implementation of: PixLoc, from the paper "Back to the Feature: Learning Robust Camera Localization from Pixels to Pose" by Paul-Edouard Sarlin*, Ajaykumar Unagar*, Måns Larsson, Hugo Germain, Carl Toft, Victor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, and Torsten Sattler.

Our repository adapts some code from S2DHM, GeM and hloc.

Authors: Javier Morlana and J.M.M. Montiel.

Related Publication:

Javier Morlana and J.M.M. Montiel, Reuse your features: unifying retrieval and feature-metric alignment, ICRA 2023. PDF

@inproceedings{morlana2023reuse,
  title={Reuse your features: unifying retrieval and feature-metric alignment},
  author={Morlana, Javier and Montiel, JMM},
  booktitle={2023 IEEE International Conference on Robotics and Automation (ICRA)},
  pages={6072--6079},
  year={2023},
  organization={IEEE}
}