Attention-based context aggregation network for monocular depth estimation

Abstract

Depth estimation is a traditional computer vision task, which plays a crucial role in understanding 3D scene geometry. Recently, algorithms that combine the multi-scale features extracted by the dilated convolution based block (atrous spatial pyramid pooling, ASPP) have gained significant improvements in depth estimation. However, the discretized and predefined dilation kernels cannot capture the continuous context information that differs in diverse scenes and easily introduce the grid artifacts. This paper proposes a novel algorithm, called attention-based context aggregation network (ACAN) for depth estimation. A supervised self-attention model is designed and utilized to adaptively learn the task-specific similarities between different pixels to model the continuous context information. Moreover, a soft ordinal inference is proposed to transform the predicted probabilities to continuous depth values which reduce the discretization error (about 1% decrease in RMSE). ACAN achieves state-of-the-art performance on public monocular depth-estimation benchmark datasets. The source code of ACAN can be found in https://github.com/miraiaroha/ACAN.