Improving Monocular Depth Estimation by Leveraging Structural Awareness and Complementary Datasets

• DOI: 10.48550/arxiv.2007.11256
• Publication Date: 2020-01-01
• Authors: Tian Chen; Shijie An; Yuan Zhang; Chongyang Ma; Huayan Wang; Xiaoyan Guo; Wen Zheng
• Keywords: Margin (machine learning); Monocular; Feature (linguistics); Benchmark (surveying)

Monocular depth estimation plays a crucial role in 3D recognition and understanding. One key limitation of existing approaches lies their lack structural information exploitation, which leads to inaccurate spatial layout, discontinuous surface, ambiguous boundaries. In this paper, we tackle problem three aspects. First, exploit the relationship visual features, propose structure-aware neural network with attention blocks. These blocks guide global structures or local details across different feature layers. Second, introduce focal relative loss for uniform point pairs enhance constraint prediction, explicitly increase penalty on errors depth-wise regions, helps preserve sharpness results. Finally, based analysis failure cases prior methods, collect new Hard Case (HC) Depth dataset challenging scenes, such as special lighting conditions, dynamic objects, tilted camera angles. The is leveraged by an informed learning curriculum that mixes training examples incrementally handle diverse data distributions. Experimental results show our method outperforms state-of-the-art large margin terms both prediction accuracy NYUDv2 generalization performance unseen datasets.