Author Type

Graduate Student

Date of Award

Fall 12-3-2025

Document Type

Thesis

Publication Status

Version of Record

Submission Date

December 2025

Department

Computer and Electrical Engineering and Computer Science

Degree Name

Master of Science (MS)

Thesis/Dissertation Advisor [Chair]

Hari Kalva

Abstract

Autonomous Vehicle (AV) systems have the potential to transform the transportation ecosystem significantly by improving safety and accessibility. One of the core components of any AV is a real-time object detection system. It helps an AV localize its position in the surrounding environment and perform tasks such as maneuver planning and obstacle detection. To make a deep-learning-based object detection system safe and robust, it must be trained on diverse and rare scenarios that can occur in road conditions. However, it is expensive and time consuming to manually create training datasets containing rare road objects or conditions. Synthetic training data can be a practical solution to this problem.

This thesis proposes a Generative AI pipeline that takes images from an existing training dataset, identifies a suitable region in each image to create an inpainting mask, places a desired rare road-scenario object via text-guided inpainting, and generates annotations of the inpainted object within the placement mask. In this work, I implemented the proposed Generative AI pipeline to inpaint vehicles on fire into road images from the NuScenes dataset and generated bounding-box annotations of the fire regions in YOLO format.

To evaluate the e!ectiveness of the synthetically generated images, I conducted an experiment by training two YOLO models from Ultralytics: one using a real world training dataset and another using synthetically generated inpainted images. Both models were tested on real world images of vehicles on fire, sourced from the internet and manually annotated by me. The YOLO model trained with synthetic images demonstrated comparable performance to the benchmark model trained with real world data, showing that synthetic images can substantially reduce the challenges of curating training datasets for rare road scenarios. Furthermore, I discussed the challenges faced while implementing the synthetic data generation pipeline, such as the complexity of identifying and annotating fire in images.

Share

COinS