Object Depth Estimation from a Single Image Using Fully Convolutional Neural Network
Ahmed J. Afifi, Olaf Hellwich · 2016
Convolutional Neural Network (CNN) has been used successfully in solving different computer vision tasks such as classification, detection, and segmentation. This paper addresses the problem of estimating object depth from a single RGB image. While stereo depth estimation is a straightforward task, predicting depth map of an object from a single RGB image is a more challenging task due to the lack of information from various image cues. Prior work focuses on exploiting geometric information or hand-crafted features. Also, they proposed the L2 norm to perform the optimization of a CNN during the training process. Using the L2 norm in regression tasks for optimization will bias the model. To address this problem, we propose a regression model with a fully convolutional neural network. To achieve robustness to outliers, we optimize the model using Tukey's biweight loss function, which is an M-estimator that is robust against outliers. The predictions are given by a single fully CNN without any post-processing techniques. In our experiments, we show that the quantitative and the qualitative results of using Tukey's biweight loss for optimization are better than of using L2 norm.