摘要
为解决目前单目图像深度估计过程中物体边界处深度跳变不明显导致的遮挡难以判别、边界处深度估计准确度较低的问题,提出了一种强化边缘的单目图像深度估计方法.采用深度估计网络输出最初预测的深度图,同时采用深度补偿网络输出应补偿深度的预测值,通过融合两组网络的输出实现对最初预测的深度图中物体边界轮廓处深度值的补偿.此外,通过设计点约束损失函数,并引入多尺度特征融合损失函数进一步提升边界处的深度估计精度.在NYU Depth v2数据集和iBims数据集上的测试实验表明本文方法能有效提升深度图中物体轮廓的清晰度,使得物体遮挡判别更加容易,可进一步提升单目图像深度估计的效果.
To solve the problem that the discrimination of object occlusion is difficult and the accuracy of depth estimation at the boundary is low caused by unobvious depth jump in monocular depth estimation,a method of monocular depth estimation with enhanced edge was proposed. Two sets of convolutional neural networks were applied to the algorithm,depth estimation network output initially predicted depth map and depth compensation network output compensate value for depth prediction,and the output of two sets of networks were fused to achieve the depth compensation of the original prediction around the object boundary.The multiscale feature fusion loss function was introduced,and the loss function of point constraint was designed to further improve the depth estimation accuracy at the boundary. Through the test and verification on NYU Depth v2 dataset and iBims dataset,the experimental results show that the method in this paper can effectively improve the clarity of the object profile in the depth map,make it easier for the object to block and judge and further enhance the effect of monocular depth estimation.
作者
王泉德
王奇坤
程凯
刘子航
WANG Quande;WANG Qikun;CHENG Kai;LIU Zihang(School of Electrical Information,Wuhan University,Wuhan 430072,China)
出处
《华中科技大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2022年第3期36-42,共7页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
基金
国家自然科学基金资助项目(61701351)。
关键词
单目图像深度估计
强化边缘
深度补偿
多尺度特征融合
点约束损失
monocular depth estimation
edge enhance
depth compensation
multi-scale feature fusion
point constraint loss
