Depth Estimation from Light Field Geometry Using Convolutional Neural Networks †

Depth estimation based on light field imaging is a new methodology that has succeeded the traditional binocular stereo matching and depth from monocular images. Significant progress has been made in light-field depth estimation. Nevertheless, the balance between computational time and the accuracy of depth estimation is still worth exploring. The geometry in light field imaging is the basis of depth estimation, and the abundant light-field data provides convenience for applying deep learning algorithms. The Epipolar Plane Image (EPI) generated from the light-field data has a line texture containing geometric information. The slope of the line is proportional to the depth of the corresponding object. Considering the light field depth estimation as a spatial density prediction task, we design a convolutional neural network (ESTNet) to estimate the accurate depth quickly. Inspired by the strong image feature extraction ability of convolutional neural networks, especially for texture images, we propose to generate EPI synthetic images from light field data as the input of ESTNet to improve the effect of feature extraction and depth estimation. The architecture of ESTNet is characterized by three input streams, encoding-decoding structure, and skipconnections. The three input streams receive horizontal EPI synthetic image (EPIh), vertical EPI synthetic image (EPIv), and central view image (CV), respectively. EPIh and EPIv contain rich texture and depth cues, while CV provides pixel position association information. ESTNet consists of two stages: encoding and decoding. The encoding stage includes several convolution modules, and correspondingly, the decoding stage embodies some transposed convolution modules. In addition to the forward propagation of the network ESTNet, some skip-connections are added between the convolution module and the corresponding transposed convolution module to fuse the shallow local and deep semantic features. ESTNet is trained on one part of a synthetic light-field dataset and then tested on another part of the synthetic light-field dataset and real light-field dataset. Ablation experiments show that our ESTNet structure is reasonable. Experiments on the synthetic light-field dataset and real light-field dataset show that our ESTNet can balance the accuracy of depth estimation and computational time.