Synthetic Dataset Generation for Offroad Navigation

Overview

Advanced RGB to thermal infrared image translation using U-Net, diffusion-based, and GAN architectures trained on Freiberg Thermal dataset.

Problem

Data scarcity in thermal domain limits autonomous vehicle perception. Need synthetic training data for depth estimation and object detection in thermal imagery.

Solution

Trained multiple generative architectures (U-Net, diffusion models, GANs) for domain adaptation. Generated synthetic datasets improving downstream thermal perception performance. Validated models for addressing data scarcity challenges.

Impact

• Domain adaptation: RGB to thermal translation
• Improved performance: Enhanced depth estimation and object detection
• Multiple architectures: U-Net, diffusion, GANs comparison
• Research contribution: Addressing thermal data scarcity