基于CBAM的深度序数回归方法

CBAM-based deep ordinal regression method

本文针对单目深度估计模型深度序数回归算法中全图像编码器易丢失较大像素值像素特征信息和位置信息的缺点,提出一种基于CBAM的深度序数回归方法。首先,将CBAM嵌入到深度序数回归算法中作为全图像编码器,依次采用通道注意力机制和空间注意力机制来捕获图像完整的特征信息和位置信息,通过获得的注意力图重新调整原始特征;其次,对像素的深度值进行离散,将深度估计重新转化为序数回归问题;最后,使用回归损失函数对网络进行训练。实验结果表明,相比于其他有监督学习、半监督学习和无监督学习的方法,该方法在KITTI数据集上取得更好的效果。

Aiming at the shortcomings of the full-image encoder in the monocular depth estimation model deep oridinal regression network that it is easy to lose the feature information and position information of the truncated length value,a CBAM-based depth ordinal regression method is proposed. First embed CBAM into depth ordinal regression algorithm as a full image encoder,and use channel attention mechanism and spatial attraction mechanism to capture the complete feature information and position information of the image,and readjust the original features through the obtained attention map;then perform the corresponding depth value Discrete,transform the depth estimation into an ordinal regression problem;finally use the regression loss function to train the network. The experimental results show that,using other supervised learning,semi-supervised learning and unsupervised learning methods,this method achieves better results on the KITTI data set.