LF-DWNet: Robust Depth Estimation Network for Light Field with Disparity Warping

Abstract

AbstractLight field (LF) images can store multi-view geometry characteristics about the observed scene, which can be helpful in depth estimation. Depth estimation has attracted much attention in recent years for its widely use in the computer vision tasks. Many approaches have been proposed to estimate the depth of LF images, including conventional methods and learning-based methods. But most of them are hard to apply to different complex situations. We propose a robust depth estimation network for LF images with disparity warping (LF-DWNet), which is robust in large disparity pixels, occlusions, and noise areas. To reduce the effect of large disparity pixels, we introduce the disparity warping processing on EPI. To extract the depth feature from warped EPI and reduce the effect of occlusions and noise areas, we design a feature extraction module based on the attention mechanism. To make full use of the depth feature our attention-based module gets, we need to guide the depth estimation by the global structure information. Besides, our LF-DWNet can integrate the depth feature from multi streams of attention-based feature extraction modules and get more credible depth map. Experiments on both synthetic and real-world datasets demonstrate the effectiveness of our method.KeywordsLight fieldDepth estimationDisparity warpingAttention mechanismGlobal integration network