Author

Ian Picklo

Date of Award

December 2024

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Electrical Engineering

First Advisor

Jeremiah Neubert

Abstract

Infrared semantic segmentation networks do not suffer the same domain shift between day and night as visual networks. As such, they have numerous applications in autonomous driving and advanced driver assistance systems which have been projected to prevent up to 62\% of driving-related deaths. That said, thermal data is notably harder to label than visual data and does not have the same representation of datasets in the open literature. This work proposes a method that allows for the quick training of IR semantic segmentation networks using knowledge distillation. Using dual-capture RGB-T imagery, and visual (RGB) network can train a student IR network without the need for costly hand-labeled ground truths. In this work, a bespoke RGB-T dataset is constructed and used for the training of IR networks with an RGB teacher. The teacher network is trained on a combination of openly available labeled data and some of our own. This was done to introduce additional classes not present. The training pipeline proposed was validated with both the bespoke dataset and openly available ones. Finally, a novel RGB-T network was proposed and tested in various configurations as part of the training pipeline. The results shown confirmed that the infrared dataset greatly increased the performance of the networks in nighttime conditionals. On openly available data, the IR network outperformed the RGB network by 23.7 \% in mIoU. While the RGB-T network outperformed openly available networks trained on the same data, it did not produce improvement in training the IR networks.

Available for download on Sunday, January 17, 2027

Share

COinS