融合边缘语义信息的单目深度估计

单目深度估计研究是许多视觉任务的基础,从图像中得到边缘清晰,细节丰富的深度图对于后续任务具有重要的作用。针对当前单目深度估计模型中不能深度融合图像语义信息以及不能较好地利用图像对象的边缘信息问题,首先构建了超像素拓扑关系图,使用图神经网络提取局部边缘信息之间的相互关系,得到以超像素为节点的拓扑关系图,其次构建了基于编解码结构的深度估计与语义分割的联合模型,通过优化联合目标函数,使模型能够融合边缘语义信息,从而提高模型提取局部结构信息的能力。通过在NYU-Depth V2 数据集中进行实验验证,结果表明模型能够构建细节丰富边缘清晰的深度图,提高了单目深度视觉估计的质量,与其他模型相比,该模型具有一定的优越性。

Integrating Spatial Semantic Information for Monocular Depth Estimation

The study of monocular depth estimation is the basis of many vision tasks. Obtaining a depth map with clear edges and rich details from an image is important for subsequent tasks. Aiming at the problem that the current monocular depth estimation model cannot deeply integrate image semantic information and can not make good use of the edge information of image objects, the superpixel topology relationship map is first constructed, and the graph neural network is used to extract the relationship between local edge information. The topological relationship graph with superpixels as nodes is obtained. Secondly, a joint model of depth estimation and semantic segmentation based on the encoder-decoder structure is constructed. By optimizing the joint objective function, the model can fuse edge semantic information, thereby improving the model"s ability to extract local structural information. Through experimental verification in the NYU-Depth V2 dataset, the results show that the model can construct a depth map with rich details and clear edges, which improves the quality of monocular depth visual estimation. Compared with other models, this model has certain advantages.