Relative order constraint for monocular depth estimation

Chunpu Liu

doi:10.1007/s10489-023-04851-7

2023-07-29

PRE

DOI：10.1007/s10489-023-04851-7

原文链接

原文求助

摘要

En 中文

Monocular depth estimation, which is playing an increasingly important role in 3D scene understanding, has been attracting increasing attention in the computer vision field in recent years. The latest monocular depth estimation methods based on deep learning have achieved significant performance by exploring various network architectures. However, compared with designing larger and more complex model architectures for monocular depth estimation, leveraging scene geometry relations to boost the performance of monocular depth estimation models has been less studied. To explore further utilization of scene geometry relations on monocular depth estimation, we propose a geometry-aware constraint that makes use of relative order information to improve the performance of monocular depth estimation models. Specifically, we first design a relative order descriptor (ROD) to construct the relative order description on single scene location. Then, based on the ROD, the relative order map (ROM) is built to represent the relative order information of the whole scene. Finally, a loss term relative order loss (ROL), which relies on ROM to supervise the training process of the monocular depth estimation model is presented. Our proposed method can help monocular depth estimation models to predict more accurate depth maps. Moreover, with the geometry constraint from our method, the monocular depth estimation model can provide prediction results where high-quality scene structure can be better preserved. We conduct extensive experiments on the popular datasets NYU Depth V2 and KITTI. The experimental results demonstrate the effectiveness of our method.

Keyword:

Depth estimation

Geometry-aware constraint

Relative order