Adaptive EPI-Matching Cost for Light Field Disparity Estimation
Tun Wang,Hao Sheng,Rongshan Chen,Ruixuan Cong,Mingyuan Zhao,Zhenglong Cui
DOI: https://doi.org/10.1109/tim.2024.3488147
IF: 5.6
2024-11-26
IEEE Transactions on Instrumentation and Measurement
Abstract:Light field (LF) technology captures information from multiple directions and angles, enabling precise disparity estimation. Recently, matching cost-based approaches have advanced rapidly and shown satisfactory results. However, these methods typically depend on fixed disparity candidates, leading to inadequate utilization of candidates and making them unsuitable for LF scenes with varying baselines. Multidirection line structures of epipolar-plane images (EPIs) associate multiple viewpoints, adaptively perceiving disparity ranges and accurately matching features in real scenes. In this article, we propose an adaptive EPI-matching cost (AEMC) for LF disparity estimation, which is proven to enhance the adaptability across datasets with varying baselines. Our approach calculates pixel-level disparity candidates to keep the predicted distribution near the ground truth (GT) and matches line structures to improve accuracy. Then, to enhance robustness during the adaptive process, we introduce an intra-EPI extraction module that dynamically establishes correlations in the local EPI while supplementing spatial information. Finally, we present a network named adaptive EPI-matching cost network (AEMCNet) for LF disparity estimation. Experimental results demonstrate that AEMCNet achieves state-of-the-art (SOTA) performance and robustness on various LF datasets with different baselines. Specifically, on the sparse LF dataset, our method reduces the mean square error (mse) by 49.6%.
engineering, electrical & electronic,instruments & instrumentation