Transmission-Guided Bayesian Generative Model for Smoke Segmentation

Smoke segmentation is essential to precisely localize wildfire so that it can be extinguished in an early phase. Although deep neural networks have achieved promising results on image tasks, they are prone overconfident for smoke due its non-rigid shape and transparent appearance. This caused by both knowledge level uncertainty limited training data accurate labeling representing the difficulty ground-truth. To effectively model two types of uncertainty, we introduce a Bayesian generative simultaneously estimate posterior distribution parameters predictions. Further, images suffer from low contrast ambiguity, inspired physics-based dehazing methods, design transmission-guided local coherence loss guide network learn pair-wise relationships based pixel distance transmission feature. promote development this field, also contribute high-quality dataset, SMOKE5K, consisting 1,400 real 4,000 synthetic with pixel-wise annotation. Experimental benchmark testing datasets illustrate our achieves predictions reliable maps ignorance about prediction. Our code dataset publicly available at: https://github.com/redlessme/Transmission-BVM.