GBNet: Gradient Boosting Network for Monocular Depth Estimation

Recently, neural networks have shown promising results in estimating depth from a single image. A large amount of per-pixel ground truth depth data is required to train the neural network in supervised learning. However, the dense depth data of ground truth is challenging to collect in realistic dynamic environments. To solve this problem, many researchers propose self- and semi-supervised learning as a credible alternative. This paper proposes a novel self- and semi-supervised monocular depth estimation method, inspired by the gradient boosting method. The existing gradient boosting method provides training to several sequential, additive, and gradual models for minimizing the error. Similarly, we design our proposed network to refine the predicted depth map sequentially and gradually generate a high-quality depth map via multi-stack CNN structures. Our method shows the state-of-the-art results for monocular depth estimation on a DDAD (Dense Depth for Autonomous Driving) dataset.