EPI-Patch Based Convolutional Neural Network for Depth Estimation on 4D Light Field

A Lightweight Depth Estimation Network for Wide-Baseline Light Fields

“…Recently, deep learning methods have gained much attention in estimating depth from light fields. Heber et al [14], Luo et al [15], Heber et al [16] and Feng et al [17] feed the input of Epipolar Plane Image (EPI) to the ConvNet where the network learns the proportional relation between the slope of the epipolar line and depth. However, this relation is hard to learn in wide-baseline light fields due to the absence of the epipolar line on the EPI.…”

“…A light field image from the camera is usually separated into the so-called subaperture images, and the baseline between sub-aperture images is very narrow. To date, traditional [2][3][4][5][6][7][8][9][10][11][12][13] and ConvNet-based [14][15][16][17][18][19][20] methods have been well studied for high performance in narrow-baseline light fields, and achieved a low percentage of errors, e.g., EPINET [19]. For wide-baseline light fields, they are usually captured by a camera array or gantry (i.e., Manuscript received 2020.…”

Manet: Multi-Scale Aggregated Network For Light Field Depth Estimation

“…Based on the epipolar plane image (EPI) or epipolar geometry property, [10] proposed to formulate the depth estimation as a classification problem, in which a standard CNN-architecture is employed on horizontal and vertical EPI patches. Since a shallow CNN is inadequate to guarantee the accuracy, a global optimization with traditional approach is utilized.…”

“…Light field depth estimation now turns into an active research topic with its potential to obtain more accurate depth maps [1][2][3][4][5][6][7][8][9][10][11][12][13][14][15][16], which is important to many depth-based vision applications, such as 3D reconstruction, semantic segmentation. In these vision fields, deep learning approaches typically surpass the traditional methods, both in accuracy and speed.…”

“…Moreover, a time-consuming global optimization is often adopted in order to improve depth accuracy [1, 3-6, 9, 11, 12]. Following the success of convolutional neural networks (ConvNets) in vision applications, some methods propose to adopt ConvNets to extract features for coarse estimates, followed by a global optimization [10,13] or refinement network [15] to improve the accuracy and/or speed. Recently, Epinet [14], a network trained end-to-end without postprocessing, achieves state-of-the-art accuracy onto the HCI and CVIA-HCI datasets [17,18].…”