Deep 3D Geometric Saliency Estimation from Light Field Images

In recent years, Light Field (LF) technology has shown remarkable progress in various computer vision tasks such as depth estimation and salient object extraction. One significant factor contributing to this advancement is the availability of commercial LF cameras that are becoming more affordable and sophisticated. Geometric saliency extraction has also played a crucial role in many computer vision problems by focusing on the most informative aspects of an image or video. In this study, we investigate the feasibility of extracting geometric salient directly from LF images. However, the task is challenging due to the lack of an extensive LF dataset that could provide rich information for image analysis. Therefore, we propose to bridge the gap by creating a synthetic dataset of LF images, which can be used for saliency map estimation. We further train a popular neural network model, called EpiNET, commonly used for depth estimation, to extract salient maps. Experimental results demonstrate that the proposed method effectively extracts salient maps with a 10-20% error on a custom metric. This finding not only confirms the feasibility of the task but also paves the way for further research in this area.