Map-based Localization for Autonomous Driving

Workshop Information

• When: August 23, 2020

• Where: Virtual: Morning Session / Afternoon Session

• Time: 08:00 - 11:00 UTC+1, 16:00 - 19:00 UTC+1

• Schedule:

  • 08:00 - 08:10: Introduction to the Workshop

  • 08:10 - 08:50: Invited Talk: Dengxin Dai

  • 08:50 - 09:30: Tech Talk: Niclas Zeller

  • 09:30 - 09:50: Contributed Paper

    • Zimin Xia, Olaf Booij, Marco Manfredi, Julian F. P. Kooij: Geographically Local Representation Learning with a Spatial Prior for Visual Localization

  • 09:50 - 10:30: Invited Talk: Andreas Geiger

• • 16:00 - 16:10: Introduction to the Workshop

  • 16:10 - 16:50: Invited Talk: Stefan Leutenegger

  • 16:50 - 17:30: Invited Talk: Raquel Urtasun

  • 17:30 - 18:10: Re-Localization Challenge

    • 17:30: 1st place: Paul-Edouard Sarlin: Hierarchical Localization with hloc and SuperGlue

    • 17:50: 2nd place: Iaroslav Melekhov: Image Stylization for Robust Features

  • 18:10 - 18:50: Invited Talk: Cyrill Stachniss

About

In this workshop, we will discuss the importance of map-based real-time localization for autonomous driving. We will focus on the problem of map generation and how to keep the maps up-to-date as well as which sensor technologies can be used. Furthermore, a new multi-season/multi-weather large-scale dataset specifically designed for the application of map-based localization in autonomous driving will be introduced. The workshop will host a Map-based Visual Re-Localization Challenge, a new challenge that evaluates the long-term performance of visual SLAM systems under changing weather conditions.

Deadlines

Paper submission due: July 08, 2020

Acceptance notification: July 14, 2020

Camera Ready: July 16, 2020

Challenge deadline: August 16, 2020

Topics

The workshop topics include (but are not limited to):

• Visual indirect/direct SLAM

• Developing visual localization methods to deal with conditions where current methods fail (weather changes, day versus night, etc.)

• Self-supervised / semi-supervised learning

• Domain adaptation

• Mapping

Detailed Description

This workshop is about map-based localization in the context of autonomous driving (AD).

By map-based localization, we understand the problem of accurately localizing (estimating the ego-position and -orientation) an autonomous vehicle in real-time in a pre-built map. Centimeter-accurate continuous global localization is a key feature for AD as it allows to position and tracks the ego-vehicle precisely within an HD map which contains important information about the environment. Being able to accurately localize within a pre-build map using standard perceptive sensors (e.g. camera, radar, LiDAR) extends the operation to GNSS-denied environments such as urban canyons or tunnels.

This task comprises several challenges including the question on how to create maps that are compressed in size and guarantee reliable localization independent of environmental conditions (e.g. weather, lighting, the season of the year) as well as keeping them up-to-date. Another aspect is the right sensor choice (with respect to robustness, accuracy, price) for both map generation and online localization.

Besides discussing the importance of map-based localization with experts from academia and industry the workshop will introduce our new dataset to the community. We will present a multi-season/multi-weather large-scale dataset specifically designed for the application of map-based localization in AD. The dataset consists of sequences from different road scenarios around the city of Munich. The same scenarios have been recorded multiple times under different weather and lighting conditions. The real-world data was collected using an Artisense VINS, which is a stereo-inertial sensor combined with RTK-GNSS. Fusing all sensor data in a direct SLAM approach gives centimeter-accurate geo-referenced 6D poses for all images in the sequences.

Sponsors