.png){kind=icon}
.png){kind=icon}
Return
Multiplane depth image for view-consistent light field depth estimation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2025.114363.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Light field cameras capture both spatial and angular information of light rays, offering rich data that has made center-view depth estimation a topic of significant interest in recent years. However, the limited flexibility of center-view depth estimation presents substantial challenges for real-world applications, underscoring the need for full-view depth estimation. A primary challenge in this domain is ensuring view consistency. Multiplane Image is a widely used technique for view synthesis, which represents a 3D scene as a series of parallel 2D image planes. Inspired by this structure, we propose a novel approach called Multiplane Depth Image (MDI), which leverages the consistency of density features across multiple views to represent depth information more effectively. To accurately capture the spatial relationships of occluded objects, we introduce a scene-wide hierarchical density update mechanism, which renders depth from the foreground to background, facilitating a more coherent depth representation. Additionally, it corrects visible holes during propagation by focusing on the evident missing regions in the updated density. Finally, we develop an LF full-view depth estimation framework based on these techniques, which enables simultaneous depth prediction across all views. This framework incorporates a comprehensive loss function to supervise depth errors, view consistency, and edge blurring. Experimental results demonstrate that our method predicts high-quality depth maps across all views and achieves state-of-the-art performance compared to center-view methods.