Encoder-decoder with densely convolutional networks for monocular depth estimation

We propose an encoder-decoder with densely convolutional networks model to recover the depth information froma single RGB image without the need for depth sensors. The encoder part serves to extract the most representative information from the original data through a series of convolution operations and to reduce the resolution of the spatial input feature. We use the decoder section to produce an upsampling structure that improves the output resolution. Our model is trained from scratch, without any special tuning process, and uses a new optimization function to adaptively learn the rate. We demonstrate the effectiveness of the method by evaluating both indoor and outdoor scenes, and the experimental results show that our proposed approach is more accurate than competing methods. (C) 2019 Optical Society of America