SCENE CLASSIFICATION FOR DEPTH ASSIGNMENT

Abstract

Due to development of 3D display technology, industries related 3D have been grown. For this reason, the demand of 3D contents increases, but there is a short sup- ply of 3D contents. Consequently, research on 2D-to-3D conversion is underway. In 2D-to-3D conversion, the depth information of scene is obtained through an analysis of several depth cues on input sequence and the depth map corresponding to a scene can be generated by combining several depth cues and assigning an appropriate depth level. Scene classification for depth assignment is needed in this process. This paper classifies a scene into landscape, linear perspective, and normal type automatically. The proposed method analyzes landscape type and found there is a relation between image pattern and distribution of color and edge, and suggest the criteria for clas- sification. Moreover, the other criteria for linear perspective classification based on vanishing point detection is proposed. To verify performance, the proposed features are fed into a linear SVM classifier, and 651 images are used. Experiment results show that the algorithm has an advantages in performance by about 13%.