Dilated Fully Convolutional Neural Network for Depth Estimation from a Single Image
Yindong Hua, Yifeng Liu, Binghan Li, Mi Lu · 2019
Convolutional Neural Network (CNN) has achieved particularly good results on depth estimation from a single image. However, certain disadvantages exist including: (1) Traditional CNNs adopt pooling layers to increase the receptive field, but it will lower the resolution and cause the information loss. (2) Almost all frameworks of CNN proposed for depth estimation apply the fully connected layers to obtain global information and they will introduce too many parameters which often result in out-of-memory problems. In this paper, we present a new module named dilated fully convolutional neural network to tackle these disadvantages. On one hand, the developed method takes advantages of dilated convolutions that can support the exponential expansion of the receptive field without loss of resolution; On the other hand, our module replaces the fully connected layers with the fully convolutional layers, which can significantly reduce the number of parameters to make our module more universal. By experiments, we show that the presented module achieves state-of-the-art results on NYU Depth V2 datasets.