A deep generative directed network for scene depth ordering

作者：Wu, Kewei^*; Gao, Yang; Ma, Hailong; Sun, Yongxuan; Yao, Tingting; Xie, Zhao

来源：Journal of Visual Communication and Image Representation, 2019, 58: 554-564.

DOI：10.1016/j.jvcir.2018.12.034

• Deep generative directed-network
• Depth ordering
• Hidden Markov field
• DenseNet

In this paper, we present a Deep Generative Directed-Network (DGDN), which estimates the occlusion relationship of boundaries. Specially, we use a low-level segmentater to partition the image into regions, then estimate their occlusion relationship by two perceptual depth cues. We decompose our DGDN model into three sub-modules to extract regional appearance cue, edgel orientation cue and to further infer global occlusion relationship with these cues, respectively. Firstly, we predict regional scene depth by a upsampling deep dense network (DenseNet). Secondly, we simultaneously estimate edgel occlusion with logistic regression. However, the occlusion relationship always suffers from unexpected conflicts due to noisy regional and edgel cues. Therefore, we finally infer occlusion relationship in a Hidden Markov Field (HMF), which tackles conflicts with bi-direction inference and the HMF parameters are exploited by iterative EM-like procedure. Ablation experiments on NYUv2 and Make3D database prove that our DGDN model outperforms state-of-the-art methods.

• 出版日期2019-1
• 单位合肥工业大学; 大连海事大学