MaSim2RealForAD

Chair for Computer Aided Medical Procedures & Augmented Reality
Lehrstuhl für Informatikanwendungen in der Medizin & Augmented Reality

THIS WEBPAGE IS DEPRECATED - please visit our new website

Semantically Consistent Sim2Real Domain Adaptation for Autonomous Driving.

Abstract

A broad variety of real-world scenarios require autonomous systems to rely on machine learning-based perception algorithms. Such algorithms are knowingly data-dependent, yet data acquisition and labeling is a costly and tedious process. One of the common alternatives to real data acquisition and annotation is represented by simulation and synthetic data. Despite being a powerful research tool, synthetic data typically reveals a significant domain gap with respect to target real data. The underlying phenomenon has been defined as a covariate shift. This problem is normally a subject for domain adaptation methods such as sim2real domain transfer.


Source: https://junyanz.github.io/CycleGAN/

Details

- You build upon state-of-the-art techniques for image domain transfer.
- You help to design a novel method for synthetic to real domain adaptation of the traffic scenes.
- You conduct experiments with different training setups using cluster.
- You evaluate the proposed approach in the task of image semantic segmentation.

Requirements

- Solid background in machine learning, experience with Pytorch, Linux.
- Excellent programming skills in Python.
- Good knowledge of software development.
- Team and communication skills.

Literature

Xun Huang, Ming-Yu Liu, Serge Belongie, Jan Kautz, "Multimodal Unsupervised Image-to-Image Translation", ECCV 2018

Students.ProjectForm
Title: Semantically Consistent Sim2Real Domain Adaptation for Autonomous Driving.
Abstract: A broad variety of real-world scenarios require autonomous systems to rely on machine learning-based perception algorithms. Such algorithms are knowingly data-dependent, yet data acquisition and labeling is a costly and tedious process. One of the common alternatives to real data acquisition and annotation is represented by simulation and synthetic data. Despite being a powerful research tool, synthetic data typically reveals a significant domain gap with respect to target real data. The underlying phenomenon has been defined as a covariate shift. This problem is normally a subject for domain adaptation methods such as sim2real domain transfer.
Student:  
Director: Federico Tombari
Supervisor: Artem Savkin
Type: Master Thesis
Area: Machine Learning, Computer Vision
Status: open
Start: 01.10.2021
Finish:  
Thesis (optional):  
Picture:  


Edit | Attach | Refresh | Diffs | More | Revision r1.2 - 28 Jul 2021 - 13:25 - ArtemSavkin