Map-free Visual Relocalization:
Metric Pose Relative to a Single Image

Eduardo Arnold · Jamie Wynn · Sara Vicente · Guillermo Garcia-Hernando · Áron Monszpart · Victor Adrian Prisacariu · Daniyar Turmukhambetov · Eric Brachmann

ECCV 2022

Project Page | Paper | arXiv | Supplemental

This is the reference implementation of the paper "Map-free Visual Relocalization: Metric Pose Relative to a Single Image" presented at ECCV 2022.

Standard visual relocalization requires hundreds of images and scale calibration to build a scene-specific 3D map. In contrast, we propose Map-free Relocalization, i.e., using only one photo of a scene to enable instant, metric scaled relocalization.

We crowd-sourced a substantial new dataset for this task, consisting of 655 places. We also define a new benchmark based on this dataset that includes a public leaderboard.

Overview

Setup
Our dataset
Evaluate your method
Baselines: Relative Pose Regression
Baselines: Feature Matching + Scale from Estimated Depth
Extended Results (7Scenes & Scannet)
Cite
License
Changelog
Acknowledgements

🔩 Setup

Using Anaconda, you can install dependencies with

conda env create -f environment.yml
conda activate mapfree

We used PyTorch 1.8, PyTorch Lightning 1.6.5, CUDA toolkit 11.1, Python 3.7.12 and Debian GNU/Linux 10.

📷 Map-free Visual Relocalization Dataset

We introduce a new dataset for development and evaluation of map-free relocalization. The dataset consists of 655 outdoor scenes, each containing a small ‘place of interest’ such as a sculpture, sign, mural, etc.

To use our code, download our dataset and extract train/val/test.zip files into data/mapfree.

Organization

The dataset is split into 460 training scenes, 65 validation scenes and 130 test scenes.

Each training scene has two sequences of images, corresponding to two different scans of the scene. We provide the absolute pose of each training image, which allows determining the relative pose between any pair of training images.

For validation and test scenes, we provide a single reference image obtained from one scan and a sequence of query images and absolute poses from a different scan.

An exemplar scene contains the following structure:

train/
├── s00000
│   ├── intrinsics.txt
│   ├── overlaps.npz
│   ├── poses.txt
│   ├── seq0
│   │   ├── frame_00000.jpg
│   │   ├── frame_00001.jpg
│   │   ├── frame_00002.jpg
│   │   ├── ...
│   │   └── frame_00579.jpg
│   └── seq1
│       ├── frame_00000.jpg
│       ├── frame_00001.jpg
│       ├── frame_00002.jpg
│       ├── ...
│       └── frame_00579.jpg

intrinsics.txt

Encodes per frame intrinsics with format

frame_path fx fy cx cy frame_width frame_height

poses.txt

Encodes per frame extrinsics with format

frame_path qw qx qy qz tx ty tz

where $q$ is the quaternion encoding rotation and $t$ is the metric translation vector.

Note:

The pose is given in world-to-camera format, i.e. $R(q), t$ transform a world point $p$ to the camera coordinate system as $Rp + t$.
For val/test scenes, the reference frame (seq0/frame_00000.jpg) always has identity pose and the pose of query frames (seq1/frame_*.jpg) are given relative to the reference frame. Thus, the absolute pose of a given query frame is equivalent to the relative pose between the reference and the query frames.
We DO NOT provide ground-truth poses for the test scenes. These are kept private for evaluation in our online benchmarking website. The poses provided for test sequences are invalid lines containing 0 for all parameters.
There might be "skipped frames", i.e. the linear id of a frame does not necessarily correspond to its frame number.

overlaps.npz

Available for training scenes only, this file provides the overlap score between any (intra- and inter-sequence) pairs of frames and can be used to select training pairs. The overlap score measures the view overlap between two frames as a ratio in the interval $[0,1]$, computed based on the SfM co-visibility. Details of how this is computed is available in the supplemental materials.

The file contains two numpy arrays:

idxs: stores the sequences and frame numbers for a pair of images (A, B), for which the overlap is computed. Format: seq_A, frame_A, seq_B, frame_B
overlaps: which gives the corresponding overlap score.

For example, to obtain the overlap score between frames seq0/frame_00023.jpg and seq1/frame_00058.jpg one would do:

f = np.load('overlaps.npz', allow_pickle=True)
idxs, overlaps = f['idxs'], f['overlaps']
filter_idx = (idxs == np.array((0, 23, 1, 58))).all(axis=1)
overlap = overlaps[filter_idx]

Note:

Although we computed overlap scores exhaustively between any two pairs, we only provide rows for pairs of frames with non-zero overlap score.

Data Loader

We provide a reference PyTorch dataloader for our dataset in lib/datasets/mapfree.py.

📊 Evaluate Your Method

We provide an online benchmark website to evaluate submissions on the test set.

Submission Format

The submission file is a ZIP file containing one txt file per scene:

submission.zip
├── pose_s00525.txt
├── pose_s00526.txt
├── pose_s00527.txt
├── pose_s00528.txt
├── ...
└── pose_s00654.txt

Each of the text files should contain the estimated pose for the query frame with the same format as poses.txt, with the additional confidence column:

frame_path qw qx qy qz tx ty tz confidence

Note that the evaluation only considers every 5th frame of the query sequence, so one does not have to compute the estimated pose for all query frames. This is accounted for in our dataloader.

Submission Script

We provide a submission script to generate submission files:

python submission.py <config file> [--checkpoint <path_to_model_checkpoint>] -o results/your_method

The resulting file results/your_method/submission.zip can be uploaded to our online benchmark website and compared against existing methods in our leaderboard.

Local evaluation

We do NOT provide ground-truth poses for the test set. But you can still evaluate your method locally, e.g. for hyperparameter tuning or model selection, by generating a submission on the validation set

python submission.py <config file> [--checkpoint <path_to_model_checkpoint>] --split val -o results/your_method

and evaluate it on the validation set using

python -m benchmark.mapfree results/your_method/submission.zip --split val

This is the same script used for evaluation in our benchmarking system, except we use the test set ground-truth poses.

Exemples of submissions for existing baselines

You can generate submissions for the Relative Pose Regression and Feature Matching baselines using

# feature matching (SuperPoint+SuperGlue), scale from depth (DPT KITTI), Essential Matrix solver
python submission.py config/matching/mapfree/sg_emat_dptkitti.yaml -o results/sg_emat_dptkitti

# feature matching (LoFTR), scale from depth (DPT NYU), PnP solver
python submission.py config/matching/mapfree/loftr_pnp_dptnyu.yaml -o results/loftr_pnp_dptnyu

# relative pose regression model, 6D rot + 3D trans parametrization
python submission.py config/regression/mapfree/rot6d_trans.yaml --checkpoint weights/mapfree/rot6d_trans.ckpt -o results/rpr_rot6d_trans

# relative pose regression model, 3D-3D correspondence parametrization + Procrustes
python submission.py config/regression/mapfree/3d3d.yaml --checkpoint weights/mapfree/3d3d.ckpt -o results/rpr_3d3d

You can explore more methods by inspecting config/matching/mapfree and config/regression/mapfree.

Relative Pose Regression Baselines

Pre-trained Models

We provide Mapfree models and Scannet models for all the RPR variants presented in the paper/supplemental. Extract all weights to weights/. The models name match the configuration files in config/regresion/

Custom Models

One can customize the existing models by changing e.g. encoder type, feature aggregation variant, output parametrisation and loss functions. All these hyper-parameters are specified in the configuration file for a given model variant. See e.g. config/regression/mapfree/3d3d.yaml.

We provide multiple variants for the encoder, aggregator and loss functions.

One can also define a custom model by registering it in lib/models/builder.py. Given a pair of RGB images, the model must be able to estimate the metric relative pose between the pair of cameras.

Training a Model

To train a model, use:

python train.py config/regression/<dataset>/{model variant}.yaml \
                config/{dataset config}.yaml \
                --experiment experiment_name

Resume training from a checkpoint by adding --resume {path_to_checkpoint}

The top five models, according to validation loss, are saved during training. Tensorboard results and checkpoints are saved into the folder weights/experiment_name.

Feature Matching + Scale from Depth Baselines

We provide different feature matching (SIFT, SuperPoint+SuperGlue, LoFTR), depth regression (DPT KITTI, NYU) and pose solver (Essential Matrix Decomposition, PnP) variants.

One can choose the different options for matching, depth and pose solvers by creating a configuration file in config/matching/mapfree/.

Download correspondences and depth files

To reproduce feature matching methods baselines

Download DPT estimated depth maps.
Download feature-matching correspondences (LoFTR and SuperPoint+SuperGlue).
Extract both files to data/mapfree

Custom feature matching method

We provide pre-computed correspondences (SIFT, SuperGlue+SuperPoint and LoFTR) in the path data/mapfree/{val|test}/{scene}/correspondences_{feature_method}.npz

To try out your own feature matching methods you need to create a npz file storing the correspondences between the reference frame and all query frames for each scene. See steps below:

Create a wrapper class to your feature matching method in etc/feature_matching_baselines/matchers.py
Add your wrapper into MATCHERS in etc/feature_matching_baselines/compute.py
Execute etc/feature_matching_baselines/compute.py using your desired feature matcher on the Mapfree dataset.
Create a new configuration file for your feature-matching baseline, e.g. modify config/matching/mapfree/sg_emat_dptkitty.yaml by replacing SG in MATCHES_FILE_PATH to the name of your matcher.

Note on recomputing SG/LoFTR correspondences

To use SG/LoFTR you need to recursively pull the git submodules using

git pull --recurse-submodules

Then,

cd etc/feature_matching_baselines
python compute.py -ds <Scannet or 7Scenes or Mapfree> -m <SIFT or SG or LoFTR>

For different 7Scenes pairs variants, include --pair_txt test_pairs_name.txt

You also need to download indoor/outdoor weights of LoFTR and extract them to etc/feature_matching_baselines/weights/.

Custom depth estimation method

We provide estimated metric depth maps in data/mapfree/{val|test}/{scene}/{seq}/frame_{framenum}.dpt{kitti|nyu}.png (see the dataset section)

To try your own depth estimation method you need to provide metric depth maps (png, encoded in millimeters) for each image the the validation/test set.

For example, data/mapfree/test/s00525/frame_00000.jpg, will have corresponding depth map data/mapfree/test/s00525/frame_00000.yourdepthmethod.png.

To use the custom depth maps, create a new config file, see e.g. config/matching/mapfree/sg_emat_dptkitty.yaml, and add the key ESTIMATED_DEPTH: 'yourdepthmethod'.

Custom pose solver

We provide three pose solvers: Essential Matrix Decomposition (with metric pose using estimated depth), Perspective-n-Point (PnP) and Procrustes (rigid body transformation given 3D-3D correspondences).

You can add your custom solver to lib/models/matching/pose_solver.py by creating a class that implements estimate_pose(keypoints0, keypoints1, data), where keypoints are the image plane coordinates of correspondences and data stores all information about the images, including estimated depth maps.

After creating your custom solver class, you need to register it in the FeatureMatchingModel.

Finally, you can use it by specifying POSE_SOLVER: 'yourposesolver' in the configuration file.

Results on Scannet & 7Scenes

See this page.

📜 Cite

Please cite our work if you find it useful or use any of our code

@inproceedings{arnold2022mapfree,
      title={Map-free Visual Relocalization: Metric Pose Relative to a Single Image},
      author={Arnold, Eduardo and Wynn, Jamie and Vicente, Sara and Garcia-Hernando, Guillermo and Monszpart, {\'{A}}ron and Prisacariu, Victor Adrian and Turmukhambetov, Daniyar and Brachmann, Eric},
      booktitle={ECCV},
      year={2022},
    }

️:page_with_curl: License

📝 Changelog

Acknowledgements

We use part of the code from different repositories. We thank the authors and maintainers of the following repositories.

weiliangxie/map-free-reloc

Map-free Visual Relocalization:Metric Pose Relative to a Single Image