ICCV 2025 Workshop on Large Scale Cross Device Localization

About the Workshop

As computer vision systems increasingly transition into real-world applications, reliable and scalable localization across heterogeneous devices becomes critical. The CroCoDL workshop brings together researchers from computer vision, robotics, and augmented reality to address the unique challenges of cross-device, multi-agent localization in complex, real-world environments. With a focus on 3D vision, visual localization, egocentric and embodied AI, and AR/VR/MR, this workshop aims to foster dialogue around bridging the gap between academic benchmarks and real-world deployment. This inaugural edition features leading experts from academia and industry and introduces CroCoDL, a new large-scale benchmark dataset capturing synchronized sensor data from smartphones, mixed-reality headsets, and legged robots across diverse environments. Through invited talks, a paper track, and an open competition, the workshop will highlight recent advances and open challenges in localization under domain shifts, sensor diversity, and dynamic scene conditions. By uniting communities working on structure-from-motion, neural rendering, and embodied AI, CroCoDL offers a platform to drive innovation toward robust, scalable localization systems capable of operating across devices, agents, and perspectives.

Challenge Winners

🥇PICO, Bytedance Inc. Team
92.62% overall recall
Meixia Lin*, Mingkai Liu*, Shuxue Peng, DiKai Fan, Shengyu Gu, Xianliang Huang, Haoyang Ye, Xiao Liu

🥈Hoang/Khang/Gabriele Team
91.05% overall recall
Huy-Hoang Bui*, Danh-Khang Cao*, Gabriele Berton*

Schedule

1PM-5PM Honolulu Time
Monday, October 20th, 2025
Room 301B

Time	Activity
13.00 - 13.15	Opening
13.15 - 13.45	Keynote by Gabriela Csurka Privacy Preserving Visual Localization
13.45 - 14.00	Introduction of CroCoDL challenge
14.00 - 14.15	Challenge Winners
14.15 - 14.45	Keynote by Ayoung Kim Bridging heterogeneous sensors for robust and generalizable localization
14.45 - 15.00	Highlighted Paper Talk
15.00 - 16.00	Posters (248-257) and Coffee Break
16.00 - 16.30	Keynote by David Caruso Current performance of visual-inertial SLAM on egocentric data
16.30 - 17.00	Keynote by Torsten Sattler Vision Localization Across Modalities
17.00 - 17.05	Closing

Invited Speakers

Dr. Ayoung Kim

Associate Professor at Seoul National University

Ayoung Kim leads the Robust Perception for Mobile Robotics Lab at Seoul National University. Her research aims to develop robust and reliable perception systems that enhance mobile robot navigation in complex and dynamic environments. She has made significant contributions to the field, particularly in visual localization, sensor fusion, and learning-based perception techniques.

Dr. Torsten Sattler

Senior Researcher at Czech Technical University in Prague

Torsten Sattler is a senior researcher at the Czech Technical University in Prague (CTU), where he leads the Spatial Intelligence group. He works towards making 3D computer vision algorithms such as 3D reconstruction and visual localization more robust and reliable through machine learning models trained on scene understanding and 3D computer vision tasks.

Dr. Gabriela Csurka

Principal Research Scientist at Naver Labs Europe

Gabriela Csurka is a principal research scientist at Naver Labs Europe. Her work bridges fundamental research and real-world applications in robotics and augmented reality. Her contributions to the field span various topics, including feature learning, visual place recognition, and cross-domain adaptation.

Dr. David Caruso

Research Scientist at Meta Reality Labs Research

David Caruso is a Research Scientist at Meta Reality Lab with 10 years of industry experience in human localization technology, including VIO and Visual SLAM, as well as non-visual approaches through multi-sensor fusion. He obtained his PhD from University Paris-Saclay where he pioneered new Visual-Inertial-Magnetic odometry technology. At Meta, he works on the core algorithms powering Project’s Aria localization stack.

Challenge

The workshop challenge is centered around a newly accepted dataset at CVPR 2025 — CroCoDL: Cross-device Collaborative Dataset for Localization (pre-rebuttal version available at link). To advance research in visual co-localization, we introduce CroCoDL, a significantly larger andmore diverse dataset and benchmark, as shown in Figure 1. CroCoDL is the first dataset to incorporate sensor recordings from both robots and mixed-reality headsets and covering a wider range of real-world environments than any existing cross-device visual localization dataset. It includes synchronized sensor streams from three primary devices: hand-held smartphones, head-mounted HoloLens 2, and the legged robot Spot.

For this challenge, we have selected two large-scale locations — Hydrology and Succulent — where we will release mapping and query splits. These splits will be used to evaluate visual localization performance in a crossdevice setup, meaning that the map is generated using data from one device, while the goal is to localize images taken by a different device within this map. The primary evaluation metric for submissions will be single-image localization recall at 50 cm and 5 degrees of pose error.

The competition will be split into two tracks:

T1. End-to-end methods. This track includes end-to-end methods such as coordinate scene regression, pose regression, neural radiance fields, gaussian splatting, feed-forward SfM.

T2. Other methods. This track includes any other approaches that do not classify in the first track. These approaches may incorporate learned components (e.g., feature extractors, sparse / dense matchers, or keypoint refinement techniques), but are not end-to-end trainable.

To be eligible for the awards, the top-performing teams will be required to submit a detailed technical report outlining their method.

Timeline for the Challenge:

Deadline: ~~6th of October 23:59 (11:59PM) AOE~~

Call for Papers

Please note that 4-page extended abstracts generally do not conflict with the dual submission policies of other conferences. In contrast, 8-page full papers, if accepted, will appear in the proceedings and are therefore subject to the dual submission policy. This means they must not be under review or accepted at another conference at the same time.

We invite 8-page full papers for inclusion in the proceedings, as well as 4-page extended abstracts. Extended abstracts may present either new or previously published work; however, they will not be included in the official proceedings.

All submissions must be anonymous and comply with the official ICCV 2025 guidelines.