Monocular Depth Estimation with Sharp Boundary

Monocular depth estimation is the basic task in computer vision. Its accuracy has tremendous improvement decade with development of deep learning. However, blurry boundary map a serious problem. Researchers find that mainly caused by two factors. First, low-level features, containing and structure information, may be lost networks during convolution process. Second, model ignores errors introduced area due to few portions whole area, backpropagation. Focusing on factors mentioned above. Two countermeasures are proposed mitigate blur Firstly, we design scene understanding module scale transform build lightweight fuse feature pyramid, which can deal loss effectively. Secondly, propose boundary-aware function pay attention effects boundary’s value. Extensive experiments show our method predict maps clearer boundaries, performance based NYU-Depth V2, SUN RGB-D, iBims-1 competitive.