Edge-based depth gradient refinement for 2D to 3D learned prior conversion

Abstract

2D-to-3D conversion is an important task for reducing the current gap between the number of 3D displays and the available 3D content. Here, we present an automatic 2D-to-3D image conversion approach based on machine learning principles. Stemming from the hypothesis that images with a similar structure have likely a similar 3D structure, the depth of a query color image is estimated using a color plus depth image dataset. Clusters with common scene structure are computed offline. Then, a matching process is performed to select the cluster centroid which is the most similar to the query image. A prior depth map is computed fusing the depth maps of the images in this cluster. Then, an edge-based post-processing stage is applied to the prior depth map estimation to enhance the final scene depth estimation. Promising results are obtained in two commonly used databases achieving a similar performance to other much complex state-of-the-art approaches.