基于自监督深度学习的全景图像深度估计研究

Research on Self-Supervised Depth Estimation for Panoramic Images

深度估计在虚拟现实、场景重建、自动驾驶和目标检测等领域发挥着重要作用。全景图像包含全向视野信息,逐渐成为深度估计领域的研究热点。但是,全景图像存在图像畸变的问题,而且深度数据采集、标注较为困难。对此,提出采用自监督方式,利用自监督深度学习算法,引入通道优化多空间融合注意力机制,增强远距离特征提取,以获取全局和局部信息。同时,引入全景感受野块,扩充感受野以获取多尺度信息。

Depth estimation plays an important role in the fields of virtual reality,scene reconstruction,autonomous driving and object detection.Panoramic image contains omnidirectional visual field information,which has gradually become a research hotspot in depth estimation field.However,the panoramic image has the problem of image distortion,and it is difficult to collect and label the depth data.To this end,by using self-supervised deep learning algorithm,channel optimization multi-space fusion attention mechanism is introduced,and remote feature extraction is enhanced to obtain global and local information.At the same time,the panoramic receptive field is introduced to expand the receptive field to obtain multi-scale information.