GIMO: Generative Image Outpainting for Early Smoke Segmentation

Sahir Shrestha, Weihao Li, Gao Zhu, Nick Barnes

Keywords: Generative Data Augmentation, Computer Vision, Diffusion Models, Outpainting, Early Smoke Segmentation, Bushfire Detection

TL;DR: Generating small smoke regions through a generative outpainting process that preserves segmentation masks.

Abstract: The critical challenge of early smoke segmentation for wildfire detection is hindered by the inherent transparency, deformable shape, and small size of nascent smoke regions. To overcome this, we propose a novel generative data augmentation framework leveraging image outpainting to simulate fixed-camera perspective transformations, effectively generating smaller smoke instances while preserving existing segmentation labels. We employ a diffusion generative model to outpaint smoke regions, enlarging real-world smoke images with synthetic domain-matched pixels. Experiments conducted on a state-of-the-art baseline demonstrate significant improvements, achieving a 3\% increase in mean Intersection over Union (mIoU) for small smoke and a 0.9\% overall mIoU boost. These results highlight the efficacy of our generative data augmentation pipeline in mitigating data scarcity, emphasising its potential for enhancing early wildfire detection and enabling timely deployment of fire services.

Submission Number: 68