Accurate Monocular Depth Estimation via Interaction of Hierarchical Features

Abstract

Monocular depth estimation is a challenging task, which assists in understanding 3D scene geometry from the same 2D scene. The ordinal-regression-method demonstrates superior performance in this issue but naive ordinal inference strategy for inferring the final depth values and naive operations to up-sample to the desired space scale introduce significant discretization errors and object boundary confusion. Firstly, we come up with a novel inference strategy to reduce the discretization errors. And then a specifically designed decoder that completes the fusion of different hierarchical features under guidance and the fusion feature reconstruction. We evaluate on a public monocular depth-estimation benchmark dataset (NYU Depth V2). The experimental results show that the method proposed outperforms other ordinal regression methods.