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

ECCV 2024 Workshop & Challenge

Leaderboard (Single Frame) Leaderboard (Multi Frame) Dataset Benchmark Submit Contact us

The Map-free Visual Relocalization workshop investigates topics related to metric visual relocalization relative to a single reference image instead of relative to a map. This problem is of major importance to many higher level applications, such as Augmented/Mixed Reality, SLAM and 3D reconstruction. It is important now, because both industry and academia are debating whether and how to build HD-maps of the world for those tasks. Our community is working to reduce the need for such maps in the first place.

We host the first Map-free Visual Relocalization Challenge 2024 competition with two tracks: map-free metric relative pose from a single image to a single image (proposed by Arnold et al. in ECCV 2022) and from a query sequence to a single image (new). While the former is a more challenging and thus interesting research topic, the latter represents a more realistic relocalization scenario, where the system making the queries may fuse information from query images and tracking poses over a short amount of time and baseline. We invite papers to be submitted to the workshop.

Schedule 🔗

📅 On Monday, 30^th September, 2024, AM [Add to Google Calendar]

Time	Event
09:05 - 09:20	Welcome introduction by Victor Adrian Prisacariu, Niantic and Oxford University
09:25 - 09:55	Jakob Engel, Meta Spatial AI for Contextual AI This talk focuses on Spatial AI for Contextual AI: How localization and 3D spatial scene understanding will enable the next generation of smart wearables and contextually grounded AI models. I will talk about recent research including map-free and vision-free localization for AR and smart glasses, as well as semantic scene- and user- understanding; as well as talk about some of the challenges that are left.
10:00 - 10:30	Eduard Trulls, Google Beyond visual positioning: how localization turned out to provide efficient training of neural, semantic maps What do you do if you built a system that incorporates the largest ground-level corpus of imagery into a large-scale localization system serving thousands of queries per second? You add overhead and semantic data and simultaneously re-think the problem from the ground up. In SNAP, localization becomes merely a side-task and the resulting maps start to encode scene semantics, motivating an entirely new set of applications and research directions. In this talk we will give a brief overview of how we moved away from further iterations on Google's Visual Positioning Service to break with the classic visual-localization approaches. SNAP is trained only using camera poses over tens of millions of StreetView images. The resulting algorithm can resolve the location of challenging image queries beyond the reach of traditional methods, outperforming the state of the art in localization by a large margin. More interestingly though, our neural maps encode not only geometry and appearance but also high-level semantics, discovered without explicit supervision. 📜 From VPS to SNAP (11MB)
10:30 - 11:00	Coffee break and poster session
11:00 - 11:45	Vincent Leroy, Naver Labs Europe and Single Frame challenge track winner talk Grounding Image Matching in 3D with MASt3R The journey from CroCo to MASt3R exemplify a significant paradigm shift in 3D vision technologies. This presentation will delve into the methodologies, innovations, and synergistic integration of these frameworks, demonstrating their impact on the field and potential future directions. The discussion aims to highlight how these advancements unify and streamline the processing of 3D visual data, offering new perspectives and capabilities in map-free visual relocalization, robotic navigation and beyond. 📜 Grounding Image Matching in 3D with MAStƎR (200 MB)
11:50 - 12:20	Torsten Sattler, CTU Prague Scene Representations for Visual Localization Visual localization is the problem of estimating the exact position and orientation from which a given image was taken. Traditionally, localization approaches either used a set of images with known camera poses or a sparse point cloud, obtained from Structure-from-Motion, to represent the scene. In recent years, the list of available scene representations has grown considerably. In this talk, we review a subset of the available representations.
12:25 - 12:55	Shubham Tulsiani, CMU Rethinking Camera Parametrization for Pose Prediction Every student of projective geometry is taught to represent camera matrices via an extrinsic and intrinsic matrix and learning-based methods that seek to predict viewpoints given a set of images typically adopt this (global) representation. In this talk, I will advocate for an over-parametrized local representation which represents cameras via rays (or endpoints) associated with each image pixel. Leveraging a diffusion based model that allows handling uncertainty, I will show that such representations are more suitable for neural learning and lead to more accurate camera prediction. 📜 Rethinking Camera Parametrization for (Sparse-view) Pose Prediction (66 MB)
12:55 - 13:00	Closing Remarks and award photos Image of the four prize cups.

Speakers 🔗

Jakob Engel, Meta

Talk title: Spatial AI for Contextual AI
Summary: This talk focuses on Spatial AI for Contextual AI: How localization and 3D spatial scene understanding will enable the next generation of smart wearables and contextually grounded AI models. I will talk about recent research including map-free and vision-free localization for AR and smart glasses, as well as semantic scene- and user- understanding; as well as talk about some of the challenges that are left.

Bio: Jakob Engel is a Director of Research at Meta Reality labs, where he is leading egocentric machine perception research as part of Meta’s Project Aria. He has 10+ years of experience working on SLAM, 3D scene understanding and user/environment interaction tracking, leading both research projects as well as shipping core localization technology into Meta’s MR and VR product lines. Dr. Engel received his Ph.D. in Computer Science at the Computer Vision Group at the Technical University of Munich in 2016, where he pioneered direct methods for SLAM through DSO and LSD-SLAM.

Vincent Leroy, Naver Labs Europe

Talk title: Grounding Image Matching in 3D with MASt3R
Slides: 📜 Grounding Image Matching in 3D with MAStƎR (200 MB)
Summary: The journey from CroCo to MASt3R exemplify a significant paradigm shift in 3D vision technologies. This presentation will delve into the methodologies, innovations, and synergistic integration of these frameworks, demonstrating their impact on the field and potential future directions. The discussion aims to highlight how these advancements unify and streamline the processing of 3D visual data, offering new perspectives and capabilities in map-free visual relocalization, robotic navigation and beyond.

Bio: Vincent is a research scientist in Geometric Deep Learning at Naver Labs Europe. He joined 5 years ago, in 2019, after completing his PhD on Multi-View Stereo Reconstruction for dynamic shapes at the INRIA Grenoble-Alpes under the supervision of E. Boyer and J-S. Franco. Other than that, he likes hiking in the mountains and finding simple solutions to complex problems. Interestingly, the latter usually comes with the former.

Torsten Sattler, CTU Prague

Talk title: Scene Representations for Visual Localization
Summary: Visual localization is the problem of estimating the exact position and orientation from which a given image was taken. Traditionally, localization approaches either used a set of images with known camera poses or a sparse point cloud, obtained from Structure-from-Motion, to represent the scene. In recent years, the list of available scene representations has grown considerably. In this talk, we review a subset of the available representations.

Bio: Torsten Sattler is a Senior Researcher at CTU. Before, he was a tenured associate professor at Chalmers University of Technology. He received a PhD in Computer Science from RWTH Aachen University, Germany, in 2014. From Dec. 2013 to Dec. 2018, he was a post-doctoral and senior researcher at ETH Zurich. Torsten has worked on feature-based localization methods [PAMI’17], long-term localization [CVPR’18, ICCV’19, ECCV’20, CVPR’21] (see also the benchmarks at visuallocalization.net), localization on mobile devices [ECCV’14, IJRR’20], and using semantic scene understanding for localization [CVPR’18, ECCV’18, ICCV’19]. Torsten has co-organized tutorials and workshops at CVPR (’14, ’15, ’17-’20), ECCV (’18, ’20), and ICCV (’17, ’19), and was / is an area chair for CVPR (’18, ’22, ’23), ICCV (’21, ’23), 3DV (’18-’21), GCPR (’19, ’21), ICRA (’19, ’20), and ECCV (’20). He was a program chair for DAGM GCPR’20, a general chair for 3DV’22, and will be a program chair for ECCV’24.

Eduard Trulls, Google

Talk title: Beyond visual positioning: how localization turned out to provide efficient training of neural, semantic maps
Slides: 📜 From VPS to SNAP (11MB)
Summary: What do you do if you built a system that incorporates the largest ground-level corpus of imagery into a large-scale localization system serving thousands of queries per second? You add overhead and semantic data and simultaneously re-think the problem from the ground up. In SNAP, localization becomes merely a side-task and the resulting maps start to encode scene semantics, motivating an entirely new set of applications and research directions. In this talk we will give a brief overview of how we moved away from further iterations on Google's Visual Positioning Service to break with the classic visual-localization approaches. SNAP is trained only using camera poses over tens of millions of StreetView images. The resulting algorithm can resolve the location of challenging image queries beyond the reach of traditional methods, outperforming the state of the art in localization by a large margin. More interestingly though, our neural maps encode not only geometry and appearance but also high-level semantics, discovered without explicit supervision.

Bio: Eduard Trulls is a Research Scientist at Google Zurich, working on Machine Learning for visual recognition. Before that he was a post-doc at the Computer Vision Lab at EPFL in Lausanne, Switzerland, working with Pascal Fua. He obtained his PhD from the Institute of Robotics in Barcelona, Spain, co-advised by Francesc Moreno and Alberto Sanfeliu. Before his PhD he worked in mobile robotics.

Shubham Tulsiani, Carnegie Mellon University

Talk title: Rethinking Camera Parametrization for Pose Prediction
Slides:📜 Rethinking Camera Parametrization for (Sparse-view) Pose Prediction (66 MB)
Summary: Every student of projective geometry is taught to represent camera matrices via an extrinsic and intrinsic matrix and learning-based methods that seek to predict viewpoints given a set of images typically adopt this (global) representation. In this talk, I will advocate for an over-parametrized local representation which represents cameras via rays (or endpoints) associated with each image pixel. Leveraging a diffusion based model that allows handling uncertainty, I will show that such representations are more suitable for neural learning and lead to more accurate camera prediction.

Bio: Shubham Tulsiani is an Assistant Professor at Carnegie Mellon University in the Robotics Institute. Prior to this, he was a research scientist at Facebook AI Research (FAIR). He received a PhD. in Computer Science from UC Berkeley in 2018 where his work was supported by the Berkeley Fellowship. He is interested in building perception systems that can infer the spatial and physical structure of the world they observe. He was the recipient of the Best Student Paper Award in CVPR 2015.

Victor Adrian Prisacariu, Niantic and Oxford University

Talk title: TBC (opening remarks)

Bio: Professor Victor Adrian Prisacariu received the Graduate degree (with first class hons.) in computer engineering from Gheorghe Asachi Technical University, Iasi, Romania, in 2008, and the D.Phil. degree in engineering science from the University of Oxford in 2012.
He continued here first as an EPSRC prize Postdoctoral Researcher, and then as a Dyson Senior Research Fellow, before being appointed an Associate Professor in 2017.
He also co-founded 6D.ai, where he built APIs to help developers augment reality in ways that users would find meaningful, useful and exciting. The 6D.ai SDK used a standard built-in smartphone camera to build a cloud-based, crowdsourced three-dimensional semantic map of the world all in real-time, in the background. 6D.ai was acquired by Niantic in March 2020. He is now Chief Scientist with Niantic.
Victor's research interests include semantic visual tracking, 3-D reconstruction, and SLAM.

Confirmed posters 🔗

Grounding Image Matching in 3D with MASt3R
Vincent Leroy, Yohann Cabon and Jérôme Revaud
Naver Labs Europe
Improving Map-Free Localization with Depth Supervision
Hitesh Jain and Sagar Verma
Granular AI
Mismatched: Evaluating the Limits of Image Matching Approaches and Benchmarks
Sierra Bonilla, Chiara Di Vece, Rema Daher, Xinwei Ju, Danail Stoyanov, Francisco Vasconcelos, Sophia Bano
University College London
NeRF-Supervised Feature Point Detection and Description
Ali Youssef, Fransisco Vasconcelos
University College London
CrossScore: Towards Multi-View Image Evaluation and Scoring
Zirui Wang, Wenjing Bian, Victor Adrian Prisacariu
University of Oxford

🏆 2024 Challenge Winners 🔗

Single Frame (Leaderboard)

🥇

Grounding Image Matching in 3D with MASt3R
"MASt3R (Ess.Mat + D.Scale)"

Vincent Leroy, Yohann Cabon and Jerome Revaud

Naver Labs Europe

https://arxiv.org/abs/2406.09756

🥈

Metric Pose Estimation Relative to Anchor Frame for Map-free Localization
"interp_metric3d_loftr_3d2d"

Lili Zhao¹, Zhili Liu², Lei Yang¹, and Meng Guo¹

¹China Mobile Research Institute,

²N/A

https://github.com/Shepherddot/An-Implementation-For-Two-View-Metric-Pose-Estimation

🥉

Map-Free Visual Relocalization Enhanced by Instance Knowledge and Depth Knowledge

Mingyu Xiao¹, Runze Chen ¹, Haiyong Luo², Fang Zhao¹, Juan Wang³, and Xue Peng Ma³

¹Beijing University of Posts and Telecommunications,

²Research Center for Ubiquitous Computing Systems, Institute of Computing Technology, Chinese Academy of Sciences

³Shouguang Cheng Zhi Feng Xing Technology Co., Ltd.

http://arxiv.org/abs/2408.13085

Improving Map-Free Localization with Depth Supervision
"Mickey Variant GT_Depth"

Hitesh Jain and Sagar Verma

Granular AI

https://hal.science/hal-04696742v1/document

Multi Frame (Leaderboard)

No accepted entries.

Prizes 🔗

$6000 in prizes will be divided between the top submissions of the two tracks.
Niantic is also seeking partners from the growing community to co-fund and co-judge the prizes.

Update Naver Labs Europe has kindly agreed to co-sponsor the challenge with $2000!

About the Challenge 🔗

We have extended the Map-free benchmark for the challenge with a sequence-based scenario, based on feedback from senior community members. Therefore, the challenge consists of two tracks:
1. The original, single query frame to single map frame task published with the ECCV 2022 paper;
2. A new task with multiple query frames (9) and their mobile device provided, metric tracking poses.

1. Single query frame to a single "map" frame

To recap, the task in the first track consists of from a single query image predict the metric relative pose to a single map image without any further auxiliary information.

2. Multiple query frames (9) to a single "map" frame

The second track is motivated by the observation that a burst of images, capturing small motion, can be recorded while staying true to the map-free scenario: No significant data capture or exploration of the environment.
At the same time, the burst of images allows the application of multi-frame depth estimation and contains strong hints about the scene scale from the IMU sensor on device.
We created a second version of the test set and leaderboard for this track.

Stay up to date! 🔗

Please register your interest here, so we can keep you notified about news and updates!

Important Dates 🔗

Challenge start

21^st May, 2024

Workshop paper and extended abstracts submission deadline

2^nd August, 2024

Workshop paper and extended abstracts camera-ready deadline

22^nd August, 2024

Challenge end

23^rd August, 2024

Challenge submission method description deadline

27^th August, 2024

Challenge winners announcement

30^th August, 2024

Workshop date (ECCV'24)

30^th September, 2024, AM

[Add to Google Calendar]

Location 🔗

🏤📍 Suite 6, South Level Mezzanine Google Maps to Suite 6, Mezzanine, South Wing

Allianz MiCo • Milano Convention Centre
Viale Eginardo - Gate 2 (Or Piazzale Carlo Magno) – Gate 16 (for those arriving from Domodossola Metro Station)
20149 Milano, Italy

Call for Papers: ECCV Map-free Visual Relocalization Workshop & Challenge 2024 🔗

We invite submissions of workshop papers and extended abstracts to the ECCV Map-free Visual Relocalization Workshop & Challenge 2024. This workshop aims to advance the field of visual relocalization without relying on pre-built maps. The following topics and related areas are of interest:

visual relocalization,
feature-matching,
pose regression (absolute and relative),
depth estimation (monocular and multi-frame),

scale estimation,
confidence and uncertainty,
structure-from-motion

Workshop paper and extended abstract submission deadline: 2^nd August, 2024. See also Important Dates.
Sign up through the contact form to stay up to date with future announcements.

🏆 Accepted Papers (pre-prints!) 🔗

Here follows a list of pre-prints of the three workshop papers accepted to the ECCV Map-free Visual Relocalization Workshop & Challenge 2024:

Submission Guidelines 🔗

Camera ready versions of accepted workshop papers should be submitted in the ECCV main conference camera ready template for your workshop papers. The template can be accessed here and here.

1. Workshop paper 🔗

Deadline: 2^nd August, 2024.
Follow the ECCV paper submission policies.
Use the same format as for submissions to the main conference: official ECCV 2024 template.
Accepted workshop papers should follow the ECCV paper chairs' guidance.
They will most likely communicate to use the official Springer ECCV 2024 template for the camera-ready version.
Evaluation on the Niantic Map-free Relocalization Dataset is encouraged but not required.
Please submit at the workshop CMT submission portal.
Use the "Workshop paper track" category under "Subject areas".

2. Extended abstract 🔗

Deadline: 2^nd August, 2024.
Extended abstracts are not subject to the ECCV rules, so they can be in any template (e.g., ECCV or CVPR).
The maximum length of extended abstracts excluding references is the equivalent of 4 pages in the official CVPR 2024 format.
Extended abstracts can include already published or accepted works, as they are meant to provide authors with the possibility to showcase their ongoing research in areas relevant to map-free relocalization.
Extended abstracts will not be included in the conference proceedings, do not count as paper publication.
However, the authors of accepted extended abstracts will get the opportunity to present their work at the workshop poster session.
Examples of extended abstracts: Accepted Extended Abstracts from HANDS @ ECCV 2022 . E.g.,
Scalable High-fidelity 3D Hand Shape Reconstruction via Graph Frequency Decomposition. Tianyu Luan, Jingjing Meng, Junsong Yuan. [pdf]
Please submit at the workshop CMT submission portal.
Use the "Extended abstract track" category under "Subject areas".

3. Method description (~technical report) 🔗

Challenge leaderboard entry deadline: 23^rd August, 2024.
Method description publication deadline: 27^th August, 2024.
The minimum required for valid competition entries.
Accepted workshop papers or accepted extended abstracts also qualify as valid method descriptions, no need to submit more.
Must be publicly available on ArXiv or equivalent.
If there are delays with an ArXiv publication, please send us proof of pending submission.
This can be a non-peer reviewed paper in any format (e.g., ECCV or CVPR). The method description has to contain sufficient detail for replication of the leaderboard entry's results.

Competition Requirements 🔗

Entries must be submitted within the competition time limits: 21^st May, 2024 - 23^rd August, 2024.
Entries in the leaderboard that were submitted before the 21^st May, 2024 will not be considered as participating in the 2024 Map-free Visual Relocalization challenge.
Entries in the leaderboard must point to a valid method description (as outlined above under "3. Method description").

FAQ 🔗

Q:: I would like to participate in the challenge, and I don't have a paper yet for my method. What should I do?
A:: You should submit a workshop paper or an extended abstract to the workshop, or make a method description available on ArXiv.
Upon acceptance, your leaderboard entry should refer to the accepted submission.
If rejected, make sure to have a method description available on ArXiv to keep your leaderboard entry valid.
Make sure your leaderboard entry dates between the competition start and end dates (21^st May, 2024 - 23^rd August, 2024).

Q:: I would like to participate in the challenge with a method I have already published. What should I do?
A:: Make sure your leaderboard entry dates between the competition start and end dates (21^st May, 2024 - 23^rd August, 2024).
Your leaderboard entry should point to a valid method description, e.g., the link to your published paper.

Q:: I would like to participate in the challenge with a method I have already published, but I've made some changes to the method specific to the challenge. What should I do?
A:: You probably should submit an extended abstract to the workshop or make a method description available on ArXiv, citing your original work.
Resubmitting your original work without sufficient amount of differences compared to the already submitted or published version might violate the ECCV policy regarding dual submissions.
Make sure your leaderboard entry dates between the competition start and end dates (21^st May, 2024 - 23^rd August, 2024).
Your leaderboard entry should refer to a valid method description, e.g., the title of the extended abstract or the link to your method description on ArXiv.

Q:: I don't want to participate in the challenge, but I have a paper that has not been published anywhere else and is not currently under review anywhere.
Can I submit it to the workshop?
A:: Yes! Submit it as a workshop paper on the workshop CMT.
We encourage you to update the paper with evaluation on the Niantic Map-free Relocalization Dataset, but it is not required.
If your method can solve the task in the challenge, you should consider also submitting it to the leaderboard.

We look forward to your contributions advancing the field of map-free visual relocalization.
For any questions or clarifications, please contact the workshop organizers.