Generation of temporally consistent depth maps using nosie removal from video Olgierd Stankiewicz and Krzysztof Wegner ⋆ Chair of Multimedia Telecommunications and Microelectronics, Pozna´ n University of Technology, Polanka 3, 60-965 Pozna´ n, Poland {ostank,kwegner}@multimedia.edu.pl http://www.multimedia.edu.pl Abstract. This paper presents a novel approach for providing depth maps that are temporally consistent. Temporal consistency is attained by noise removal from video. Presented approach was evaluated with use of a simple noise reduction technique and state-of-the-are depth estimation algorithm. Experiments on depth-based synthesis of standard multi-view test video sequences have been performed and yielded both subjective and objective results. These results provide evidence that the proposed approach increase temporal consistency of estimated depth maps. Key words: depth map estimation, temporal consistency, temporal noise removal 1 Introduction Depth map estimation is a technology that provides 3D representation of the scene [1]. Common approach to obtaining depth data is algorithmic estimation from video. Although many such algorithms are known in literature [2], depth estimation is still a challenge, even for the most advanced state-of-the-art tech- niques. One of the biggest of challenges in this research area, is how to provide depth maps that are consistent in time. Typically, depth data for video is esti- mated independently for each frame of the sequence. Unfortunately, estimation that is independent in time, causes depth of objects in the scene to fluctuate, due to noise. Such fluctuations are adverse, because they lead to occurrence of artificial movement in 3D representation. Desired depth map temporal consis- tency means that changes of the depth of objects in time are correlated with actual motion of the objects and do not vary from frame to frame in a random way. Majority of state-of-the-art techniques that tackle temporal consistency, in various ways expand depth estimation algorithms into time domain. For exam- ple, in [3] authors propose to extend standard 4-neighborhood belief propaga- tion depth map estimation scheme [4] to 6-neighborhood scheme by addition of ⋆ This work was supported by the public funds.