dgn-pytorch

Code for the paper: Depth Generation Network (DGN): Estimating Real World Depth from Stereo and Depth Images

Introduction

Humans can perceive the depth of a scene even from a photo of it by leveraging prior knowledge and infer the unseen area by reasoning from the adjacent, or even distant part of the scene. Target at such ability, we design a DGN to incorporate 3D senses to predict the depth by inferring to a variational Bayesian model contains a interpreter network and a generation network. We maximizing the likelihood of the distribution over the generated depth, and given the synthetic ground truth, we follow the Expectation-Maximization scheme to organize and optimize the proposed model. The resulting model can estimate the relative depth of the scene with a pair of uncalibrated stereo images.

Cite

If you referred or used DGN in your article, please consider citing:

@inproceedings{dong2019depth,
  title={Depth Generation Network: Estimating Real World Depth from Stereo and Depth Images},
  author={Dong, Zhipeng and Gao, Yi and Ren, Qinyuan and Yan, Yunhui and Chen, Fei},
  booktitle={2019 International Conference on Robotics and Automation (ICRA)},
  pages={7201--7206},
  year={2019},
  organization={IEEE}
}