基于DSPP的单目图像深度估计

Monocular depth estimation based on DSPP

空洞空间金字塔池化(ASPP)在深度学习各种任务中均有应用,传统ASPP模块只考虑了提升卷积感受视野,但ASPP中的每次空洞卷积选取的像素点分散,会丢失大量像素点间的信息,而深度估计属于密集预测任务。针对ASPP模块这一弊端提出了一种动态密集的DSPP模块。该模块用一种动态卷积代替空洞卷积,结合ASPP的思想,采用不同大小的卷积尺寸,并结合通道注意力充分利用每一层的特征,解决了ASPP丢失信息的问题,与ASPP相比在大大减小模块参数量的前提下,提升了整体模型的准确率。在NYU Depth v2数据集上与主流算法相比,深度图在均方根误差(RMSE)上降低了12.5%,到0.407,并且准确率(δ<1.25)提高了3.4%,达到0.875,验证了算法的有效性。

Atrous spatial pyramid pooling(ASPP)has applications in various tasks of deep learning.The traditional ASPP module only considers improving the receptive field of convolution,but the pixels selected by each atrous convolution in ASPP are scattered,and a lot of information between points will be lost,while depth estimation is a dense prediction task.Aiming at the drawback of the ASPP module,this paper proposed a dynamic and dense DSPP module,this module replaced atrous convolution with a dynamic convolution,combined the idea of ASPP,adopted different kernel sizes,and combined channel attention to make full use of the features of each layer.It solved the problem of ASPP losing information,and improved the accuracy of the overall model on the premise of greatly reducing the amount of module parameters compared with ASPP.Compared with the mainstream algorithm on the NYU Depth v2 dataset,the depth map reduces the root mean square error(RMSE)by 12.5%to 0.407,and the accuracy(δ<1.25)increases by 3.4%to 0.875,which verifies the effectiveness of the algorithm.