Bachelor and Master Theses

To apply for conducting this thesis, please contact the thesis supervisor(s).
Title: Contrastive Self-Supervised Learning in Lane Detection application
Subject: Computer science
Level: Advanced
Description:

(This is a collaborative thesis between MDU and Zenseact)

 

Summary 

In this thesis, we aim to improve the accuracy of 3D lane detection approaches such as 3DLaneNet[1] and GenLaneNet[2] by employing a contrastive self-supervised learning approach. 

Background 

Although Deep supervised learning methods succeed, they are expensive to label data and susceptible to human error. Unsupervised learning approaches are proposed to solve the challenges associated with supervised learning[3]. Self-supervised learning is one of the newest approaches within the unsupervised learning paradigm. Self-supervised representation learning uses input data as its supervision and is advantageous for nearly all types of downstream tasks[4]. Contrastive learning is a self-supervised learning technique that aims to keep embedding augmentation versions of the same sample close together while attempting to push embeddings from different samples apart. To increase the accuracy, many researchers[1, 2] pre-train their proposal networks on a general classification dataset such as ImageNet[5]. Using pre-trained weight enhances feature extraction performance as a result of training the network with a greater variety of shapes and other visual features. Caron et al. [6] demonstrate that pre-training downstream tasks, such as object detection, through contrastive self-supervised learning requires less labeled data and improves accuracy. 

Goals and Objectives 

We will investigate the following activities in this study: 

  • Gather a massive set of unlabelled lane images to pre-train the networks 

  • Pre-train the mentioned networks by the gathered dataset using the state-of-the-art self-supervised contrastive learning method[6] 

  • Do a piece of study on different augmentation methods to find the most significant augmentation for 3D lane detection self-supervised approach 

 

 

1. Garnett, N., et al. 3d-lanenet: end-to-end 3d multiple lane detection. in Proceedings of the IEEE/CVF International Conference on Computer Vision. 2019. 

2. Guo, Y., et al. Gen-lanenet: A generalized and scalable approach for 3d lane detection. in European Conference on Computer Vision. 2020. Springer. 

3. Erhan, D., et al. Why does unsupervised pre-training help deep learning? in Proceedings of the thirteenth international conference on artificial intelligence and statistics. 2010. JMLR Workshop and Conference Proceedings. 

4. Liu, X., et al., Self-supervised learning: Generative or contrastive. IEEE Transactions on Knowledge and Data Engineering, 2021. 

5. Krizhevsky, A., I. Sutskever, and G.E. Hinton, Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 2012. 25. 

Start date:
End date:
Prerequisites:
  • Python programming 
  • Deep Learning
  • Computer Vision
IDT supervisors: Ali Zoljodi
Examiner: Masoud Daneshtalab
Comments:
Company contact: