UTFormer: A UAV Sensing-based Triple Attention Transformer Network for Semantic Segmentation

Abstract Compared to the common urban landscape semantic segmentation, with the impact of flight altitude, UAV image semantic segmentation is more challenging because small targets have a very low percentage of pixels and multi-scale features. Yet, the commonly used successive grid downsampling strategy in the current transformer-based methods omits features of small targets. Furthermore, it leads to even worse results due to the complex background interference. In reaction to this, existing strategies aim to maintain superior resolution. Nevertheless, the application of this method incurs significantly high computational costs which brings challenges to the application of assembling UAVs. So it is significant to design a novel framework to balance retaining more pixels representing small objects during downsampling and fewer computational costs. For this, we propose a UAV Sensing-based Triple Attention Transformer Network (UTFormer), which reduces complexity and enhances the segmentation of small targets simultaneously. During the overlap embedding of feature maps, the model utilizes channel attention to allocate half of the channels for global attention while adaptively selecting important semantic information from the remaining for local attention. So this triple attention strategy not only optimizes computational resources but also improves the model's ability to handle the diverse challenges presented by UAV imagery, ensuring accurate segmentation across varied object scales and densities. Our method shows superior performance on two public datasets: AeroScapes and Vaihingen, achieving 75.57\% and 78.93\% mean intersection over union. The proposed UTFormer has been released on Github:https://github.com/leoda1/UTFormer.