FY27 Master Thesis Project- Foundation models for pretraining ADAS perception, 6 months, Linköping - Sweden

Technology, Data & Digital · Data, AI & Analytics · Machine Learning · Data Science · Artificial Intelligence

In short

This Master Thesis Project focuses on investigating foundation models for pretraining ADAS perception systems and exploring self-supervised pretraining strategies for object detection with limited labels. The project involves experimental evaluation and analysis to provide recommendations for future automotive perception systems.

Responsibilities

Investigate foundation models, encoder architectures, and pretraining strategies for ADAS perception.
Develop and evaluate knowledge distillation methods for transferring knowledge from foundation models to compact ADAS perception models.
Compare different teacher models, including image-domain and multi-view spatiotemporal approaches.
Evaluate the impact of distillation on downstream tasks like lane detection, traffic sign recognition, and vehicle detection.
Investigate self-supervised pretraining strategies for 3D object detection to reduce dependence on manual annotation.
Evaluate latent-space prediction on large unlabeled datasets and pretrained encoders on 3D object detection with varying label fractions.

Requirements

4th year student on an M.Sc. program in Computer Science and Engineering, Applied Physics and Electrical Engineering, Information Technology or related field.
Specialization in machine learning and/or computer vision is preferred.
Good, documented knowledge of deep learning techniques.
Good programming skills in Python.
Machine Learning and Deep Learning knowledge.
Computer Vision knowledge.
Python and PyTorch proficiency.
Interest in autonomous driving and ADAS technologies.

Desired Qualifications

Research questions on which ADAS perception tasks benefit most from foundation model pretraining.
Research questions on how different foundation model architectures compare in terms of performance and data efficiency.
Research questions on whether a foundation model pre-training based on a 2D image task benefits a 3D application.
Research questions on how well pretrained representations transfer across different perception tasks and datasets.
Research questions on which self-supervised pretraining strategy produces the most label-efficient BEV representation for object detection.
Research questions on whether it is more beneficial to learn representations in image space or in scene-centric BEV space.
Research questions on whether BEV-level self-supervised pretraining reduces sensitivity to annotation noise.

Benefits

Thesis project for 6 months.
Opportunity to work with cutting-edge AI technologies in the automotive sector.
Experimental evaluation and quantitative analysis of models and strategies.

Skills

Machine LearningDeep LearningComputer VisionPythonPyTorch

#AI#foundation models#ADAS#perception#computer vision#deep learning#machine learning#self-supervised learning#object detection#LiDAR#camera data#Python#PyTorch