Building extraction and height estimation are two important basic tasks in remote sensing image interpretation, which are widely used in urban planning, real-world 3D construction, and other fields. Most of the existing research regards the two tasks as independent studies. Therefore the height information cannot be fully used to improve the accuracy of building extraction and vice versa. In this work, we combine the individuaL buIlding extraction and heiGHt estimation through a unified multiTask learning network (LIGHT) for the first time, which simultaneously outputs a height map, bounding boxes, and a segmentation mask map of buildings. Specifically, LIGHT consists of an instance segmentation branch and a height estimation branch. In particular, so as to effectively unify multi-scale feature branches and alleviate feature spans between branches, we propose a Gated Cross Task Interaction (GCTI) module that can efficiently perform feature interaction between branches. Experiments on the DFC2023 dataset show that our LIGHT can achieve superior performance, and our GCTI module with ResNet101 as the backbone can significantly improve the performance of multitask learning by 2.8% AP50 and 6.5% delta1, respectively.
翻译:建筑物提取和高度估计是遥感图像解译中的两个重要基础任务,广泛应用于城市规划、实际3D建设等领域。现有研究大多视两个任务为独立研究,因此高度信息无法充分利用以提高建筑物提取的准确性,反之亦然。在本文中,我们首次通过统一多任务学习网络(LIGHT)将单个建筑物的提取和高度估计相结合,同时输出建筑物的高度图、包围框和分割掩模图。具体而言, LIGHT由一个实例分割分支和一个高度估计分支组成。特别的,为了有效统一多尺度特征分支并缓解分支之间的特征跨度,我们提出了一个Gated Cross Task Interaction(GCTI)模块,可以在分支之间有效地执行特征交互。在DFC2023数据集上的实验证明,我们的LIGHT可以实现卓越的性能,并且我们的GCTI模块以ResNet101为骨干可以显著提高多任务学习的性能,分别达到2.8%AP50和6.5%delta1。