Extracting dense depth from a single image is an important and challenging task in computer vision. Recently, significant progress has been made in monocular depth estimation (MDE) by formulating this problem as a multi-label classification instead of traditional regression. However, to discretize depth into hundreds of labels results in complicate network output and a large number of parameters. In this paper, we propose to encode the discretized depth into a binary code and formulate MDE as a hierarchical binary classification (HBC) problem to reduce the complexity. Besides, by studying the distribution of individual bit of the encoded depth codes, we find that our encoding scheme can also solve the problem of depth data imbalance. We conduct experiments on KITTI and NYU Depth V2 datasets, which show that our simplified approach still achieves a comparable performance with state-of-the-art methods.
Paper
Full text
Hierarchical Binary Classification for Monocular Depth Estimation
Semantic Scholar · Computer Science · 2019
Abstract
Extracting dense depth from a single image is an important and challenging task in computer vision. Recently, significant progress has been made in monocular depth estimation (MDE) by formulating this problem as a multi-label classification instead of traditional regression. However, to discretize depth into hundreds of labels results in complicate network output and a large number of parameters. In this paper, we propose to encode the discretized depth into a binary code and formulate MDE as a hierarchical binary classification (HBC) problem to reduce the complexity. Besides, by studying the distribution of individual bit of the encoded depth codes, we find that our encoding scheme can also solve the problem of depth data imbalance. We conduct experiments on KITTI and NYU Depth V2 datasets, which show that our simplified approach still achieves a comparable performance with state-of-the-art methods.