Monocular depth estimation with hierarchical fusion of dilated CNNs and soft-weighted-sum inference
作者：
Highlights：
• We propose a deep end-to-end learning framework to monocular depth estimation by recasting it as a multi-category classification task, where both dilated convolution and hierarchical feature fusion are used to learn the scale-aware depth cues.
• Our network is able to output the probability distribution among different depth labels. We propose a soft-weighted-sum inference, which could reduce the influence of quantization error and improve the robustness.
• Our method achieves the state-of-the-art performance on both indoor and outdoor benchmarking datasets, Make3D, NYU V2 and KITTI dataset.
摘要
•We propose a deep end-to-end learning framework to monocular depth estimation by recasting it as a multi-category classification task, where both dilated convolution and hierarchical feature fusion are used to learn the scale-aware depth cues.•Our network is able to output the probability distribution among different depth labels. We propose a soft-weighted-sum inference, which could reduce the influence of quantization error and improve the robustness.•Our method achieves the state-of-the-art performance on both indoor and outdoor benchmarking datasets, Make3D, NYU V2 and KITTI dataset.
论文关键词：Monocular depth estimation,Deep convolutional neural network,Soft-weighted-sum-inference,Dilated convolution
论文评审过程：Received 2 October 2017, Revised 12 May 2018, Accepted 31 May 2018, Available online 6 June 2018, Version of Record 20 June 2018.
论文官网地址：https://doi.org/10.1016/j.patcog.2018.05.029
