Domain adaptive depth completion via spatial-error consistency

Abstract

In this paper, we introduce a novel training framework designed to address the challenge of unsupervised domain adaptation (UDA) in depth completion. Our framework aims to bridge the gap between lidar and image data by establishing a shared domain, which is a collection of the confidence of the network’s prediction. By indirectly adapting the depth network through this common domain, the problem is decomposed into two key tasks: (1) constructing the common domain and (2) adapting the depth network using the common domain. For the construction of the common domain, errors in the network’s predictions are modelled as confidence, which serves as supervision for a sub-module called the Depth Completion Plugin (DCPlugin). The purpose of the DCPlugin is to generate the confidence associated with any given dense depth prediction. To adapt the depth network using the common domain, a confidence-aware co-training task is employed, leveraging the confidence map provided by the well-adapted DCPlugin. To assess the effectiveness of our proposed approach, we conduct experiments on multiple depth networks under adaptation scenarios, namely CARLA $\to$ KITTI and VKITTI $\to$ KITTI. The results demonstrate that our method surpasses other domain adaptation (DA) techniques, achieving state-of-the-art performance. Given the limited existing work in this domain, we provide comprehensive discussions to guide future researchers in this field.