Abstract:

Due to the occlusion and motion between objects, the estimated depth maps will be blurred and appeared boundary artifacts using current self-supervised monocular depth estimation methods. To address the above problems, an anti-occlusion monocular depth estimation algorithm is proposed by designing the loss function. The proposed algorithm, ignoring the occluded pixels with higher loss, uses the minimized photometric re-projection function to match the minimum error between two adjacent frames of the target image. Moreover, the automatic masking loss is used to process the boundary artifacts caused by object movement. Finally, the comparison of experimental results on the KITTI dataset shows that the depth maps estimated by the proposed algorithm are clearer, and the boundary artifacts in these depth maps are also reduced.

Key words: self-supervised, monocular depth estimation, boundary artifacts, minimization of photometric reprojection, automatic masking loss

摘要：

目前利用自监督单目深度估计方法对城市街道进行深度估计时，由于物体间存在遮挡和运动，导致估计的深度图结果模糊以及出现边界伪影。针对上述问题，通过对损失函数进行设计，提出了一种抗遮挡的单目深度估计方法。该方法采用最小化光度重投影函数，对目标图像前后帧中选择最小误差进行匹配，忽略掉损失较高的被遮挡像素，同时采用自动掩蔽损失来处理物体运动造成的边界伪影。在KITTI数据集上的对比实验结果表明，所提方法估计的深度图结果更加清晰，并能有效减少深度图中的边界伪影。

关键词: 自监督, 单目深度估计, 边界伪影, 最小化光度重投影函数, 自动掩蔽损失