Semi-supervised Depth Estimation from Sparse Depth and a Single Image for Dense Map Construction

For robust navigation, the objective in visual SLAM is to create a dense map from a sparse input. Although there have been a significant number of endeavors on real-time mapping, the existing works for visual SLAM systems still fail to preserve adequate geometry details that are important for navigation. This paper estimates pixel-wise depth from a single image and a few depth points which are constructed from registered LiDAR or acquired from visual SLAM systems to construct a dense map. The main idea is to employ a set of new loss functions consisting of photometric reconstruction consistency (forward-backward consistency and left-right consistency), depth loss, nearby frame geometric consistency, and smoothness loss and propose a depth estimation network based on ResNet. The experimental results show that the proposed method is superior the state-of-the-art methods on both raw LiDAR scans dataset and semi-dense annotation dataset. Furthermore, the errors of the sparse depth produced by stereo ORB-SLAM2 are evaluated and this sparse depth and a single image are fed into the proposed model to further demonstrate the superiority of the proposed work.