Master thesis: Neural representations for self-supervised monocular 3D reconstruction in Lindau bei DENSO ADAS Engineering Services GmbH

DENSO ADAS Engineering Services GmbH, as a part of DENSO Mobility, is a R&D company with activities in developing the next generation of 3D digital perception based mainly on computer vision, radar and sonar technology for future driving and safety functions. As a full subsidiary of the DENSO Corporation, the world-wide second largest tier-one automotive supplier, we combine the strength of a globally operating cooperation with the agility and familiar atmosphere of a small business.

Thesis Motivation
Occupancy Networks are neural networks that predict the occupancy of voxels in a pre-defined 3D space, based on input conditions such as RGB images or other sensor information. For a given instance of input, corresponding dense ground truth semantic voxels are collected and a network is then trained to translate the input(s) to a 3D volume representing occupancy in a supervised manner. There is a growing interest in using these networks for autonomous driving perception, but this requires dense ground truth semantic voxels, which can be difficult to obtain. These are generally accumulated from human labelled semantic point clouds and require accurate ego-motion and consistent labelling. Neural Radiance Fields are another type of network representations of 3D scenes. They approximate the rendering function that takes in the viewing-direction and 3D coordinates to output color and density along the viewing ray within a pre-defined volume. While works such as Block-NeRF demonstrate their ability for large scale reconstructions, their use is motivated by offline data generation, rather than online 3D reconstruction. Furthermore, converting NeRFs to representations useful for driving perception is highly error-prone due to topological differences. Recently methods such as SceneRF and BTS have been proposed for use in driving scenarios. This thesis would propose similar representations that can use used to train 3D reconstruction networks in a self-supervised manner, that can later be deployed on the car. The aim of this thesis is to combine ideas from Occupancy Networks and recent advances in NeRFs to come up with representations useful for online reconstructions without heavily relying on accurate, expensive ground truth labels.

Main goal is to propose a novel representation for self-supervised monocular 3D reconstruction use-case in autonomous driving scenarios.

Specific tasks include:

• Review existing representations such as NeRF-based, Occupancy Networks, SDF, depth map, etc. for pros and cons to be used in driving scenarios for 3D reconstruction
• Outline data requirements and existing suitable datasets or create new ones
• Propose modifications and experiment their plausibility regarding the main goal
• Run extensive comparisons with existing representations and ablation studies
• Results analysis with a clear understanding of advantages, limitations and future work
• Analyze the usability of the proposed solution for DNAE’s purposes
• Document a good master thesis report and check possibility of publication

• Background knowledge (via relevant courses) on the topic of thesis and interest
• Ability to research, experiment and execute independently
• Good communication skills (verbal and written - great if you have published articles in conferences)
• Ability to thrive in result oriented, fast-paced, focused and ambitious work environment
• Motivated to learn new things
• Enjoy work in international teams

• Flexible working time model
• Modern offices, free parking garage and a nice outdoor area
• Creativity room for leisure breaks or brainstorming meetings
• Company and team building events (Summer Barbeque, Christmas party,...)

It has come to our attention that various individuals have contacted people offering false employment opportunities with DENSO. Such scams are fraudulent and intended to steal from the victims.

By making you aware of this, we hope to avoid and ultimately stop victims from falling for this scam. Please do not provide any personal or financial information and do not send any money to anyone you suspect of recruitment fraud.

If you are contacted by someone whom you suspect may not be appropriately representing DENSO, please email: EU_DIEU_TALENTACQUISITION@eu.denso.com. Please forward the original email that you received that includes the original subject line and complete header information.

Do not send CVs or vacancy applications to that email address as it will not be reviewed or considered. If you wish to apply for a job, please submit your application for an open position to https://careers.smartrecruiter