Postdoc in deep learning for long-term visual localization

Information about the research project
Autonomous cars need to localize themselves reliably in the road network in order to be able to plan ahead and safely follow planned paths. Localization is typically performed via a pre-built map, where the car establishes correspondences between data captured by its sensors, such as cameras, lidars and radars, and the map. These correspondences are then used to estimate the position and orientation of the car with respect to the map.

One key challenge for localization approaches, which is highly relevant to autonomous vehicles, is long-term operation. As the appearance and geometry of the scene change over time, it becomes harder and harder to detect the same landmarks in our current sensor data that we describe in our map and correctly associate them to each other. Examples for such changes include illumination changes over the course of a day or seasonal changes that affect the vegetation. However, current state-of-the-art approaches for keypoint detection and description are only moderately robust to such changes.

This project aims to tackle the long-term operation problem through continuous map updates and keypoint detector and descriptor learning. The map update step will provide training data for the detector and descriptor learning stage. Improved local features will in turn lead to better maps and more accurate localization results. To facilitate this research, we have a large image dataset with ground truth poses collected over a whole year as well as the ability to collect our own georeferenced data using Chalmers' ReVeRe-lab. Applicants are also encouraged and provided freedom to pursue their own research ideas, both within the context of the project as well as through collaborations outside of it.

Information about the research groups
The Signal Processing group conducts research in the field of physical and statistical signal and image modeling and inference. We actively pursue research in target tracking, array signal processing, estimation, detection and machine learning. Projects range from development of mathematical theory, method development and applications in the area of perception for autonomous vehicles, radar systems and biomedical devices. 

The Computer Vision Group conducts research in the field of automatic image interpretation. The group targets both medical applications, such as the development of new and more effective methods and systems for analysis, support and diagnostics, as well as general computer vision applications including autonomously guided vehicles, image-based localization, stereo and structure-from-motion. The main research problems include mathematical theory, algorithms and machine learning for inverse problems such as reconstruction, segmentation and registration.

Information about Chalmers AI Research Centre (CHAIR)
This project is funded by the Chalmers AI Research Centre (CHAIR). CHAIR aims at increasing Chalmers’ expertise and excellence in Artificial Intelligence. Its objective is to conduct world-leading research for the benefit of both industry and the public sector. CHAIR’s focus is on developing unique AI expertise in research, education, and innovation, and offering a highly attractive environment for world-leading AI researchers. The goal is to become the preferred AI partner for Swedish industry.

Major responsibilities
This is a one-year postdoc, with the ambition of extending it for a second year. The major responsibility is to perform your own research in a research group and lead the efforts on the above mentioned research project. The position may also include teaching on master's level as well as supervising master's and/or PhD students to a certain extent. Another important aspect involves collaboration within academia and with society at large. The position is meritorious for future research duties within academia as well as industry/the public sector.

Qualifications
The candidate should have a strong background in machine learning, with a focus of (deep) machine learning for computer vision. Experience in learning novel keypoint detectors or descriptors is preferred. An additional background in 3D computer vision, especially Structure-from-Motion, SLAM, or visual localization, is desirable, but not necessary.

The candidate should have published at least one paper in one of the main international conferences in the fields of computer vision (ICCV, ECCV, CVPR), robotics (ICRA, IROS) or machine learning (NeurIPS, ICML, ICLR).

To qualify for the position of postdoc, you must have a doctoral degree in a relevant field; the degree should generally not be older than three years. You are expected to be somewhat accustomed to teaching, and to demonstrate good potential within research and education. If you obtained your PhD degree from Chalmers, you must have worked as a reasearcher somewhere else for at least one year after obtaining your PhD.

The position requires sound verbal and written communication skills in English. If Swedish is not your native language, Chalmers offers Swedish courses.

Application deadline: 29 November, 2020

Please read more at Chalmers webpage.

For questions, please contact:
Lars Hammarstrand, lars.hammarstrand@chalmers.se, tel: +46317721788
Torsten Sattler, torsat@chalmers.se, tel: +46317726618

For more general questions about workning at Chalmers, please contact: Tomas McKelvey, tomas.mckelvey@chalmers.se, Tel: +46317728061