Noise-to-Dataset: A Diffusion-Based Framework for Semantic Segmentation Dataset Generation

This paper proposes a novel synthetic dataset generation framework called Noise-to-Dataset to address data scarcity in semantic segmentation tasks on the LWIR domain. The framework consists of two stages: a denoising diffusion probabilistic model (DDPM) that generates semantic masks from Gaussian noise and a semantic diffusion model (SDM) that produces synthetic images based on these masks. Noise-to-Dataset enables the creation of diverse, high-quality synthetic datasets, significantly improving segmentation model performance. Experimental results show enhancements not only in LWIR datasets but also in RGB datasets like Cityscapes and ADE20K, highlighting its potential to generate valuable training data without the need for manual annotation.