Leveraging ground-annotated data for scene analysis on unmanned aerial vehicles (UAVs) can lead to valuable real-world applications. However, existing unsupervised domain adaptive (UDA) methods primarily focus on domain confusion, which raises conflicts among training data if there is a huge domain shift caused by variations in observation perspectives or locations. To illustrate this problem, we present a ground-to-UAV fire segmentation method as a novel benchmark to verify typical UDA methods, and propose an effective framework, Colour-Mix, to boost the performance of the segmentation method equivalent to the fully supervised level. First, we identify domain-invariant fire features by deriving fire-discriminating components (u*VS) defined in colour spaces Lu*v*, YUV, and HSV. Notably, we devise criteria to combine components that are beneficial for integrating colour signals into deep-learning training, thus significantly improving the generalisation abilities of the framework without resorting to UDA techniques. Second, we perform class-specific mixing to eliminate irrelevant background content on the ground scenario and enrich annotated fire samples for the UAV imagery. Third, we propose to disentangle the feature encoding for different domains and use class-specific mixing as robust training signals for the target domain. The framework is validated on the drone-captured dataset, Flame, by using the combined ground-level source datasets, Street Fire and Corsica Wildfires. The code is available at https://github.com/Rui-Zhou-2/Colour-Mix.
Read full abstract