EPI-Guided Cost Construction Network for Light Field Disparity Estimation

Tun Wang, Rongshan Chen, Ruixuan Cong, Da Yang, Zhenglong Cui, Fangping Li, Hao Sheng; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2023, pp. 3438-3446

Abstract

Recent learning-based light field (LF) disparity estimation methods construct cost volume by sequentially shifting each sub-aperture image (SAI) with a series of predefined offsets. They only use the visual information of SAIs and lose the geometry of LF. In this paper, we design a simple network that can cleverly integrate EPI features with cost volume to estimate the disparity. Firstly, we propose an efficient EPI extraction module to use abundant line characteristics. Secondly, we offer an EPI-Cost volume construction module that can create volume guided by the EPI line and the color consistency of images. Finally, after completing it, we adopt an intervolume fusion module to considerably correlate the validity of EPI lines in both directions. Experimental results show the proposed method achieves state-of-the-art performance in the quantitative and qualitative evaluation of the UrbanLF-Syn dataset.

Related Material

[pdf]
[bibtex]
@InProceedings{Wang_2023_CVPR, author = {Wang, Tun and Chen, Rongshan and Cong, Ruixuan and Yang, Da and Cui, Zhenglong and Li, Fangping and Sheng, Hao}, title = {EPI-Guided Cost Construction Network for Light Field Disparity Estimation}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops}, month = {June}, year = {2023}, pages = {3438-3446} }