TY - GEN
T1 - Effective removal of artifacts from views synthesized using Depth Image Based Rendering
AU - Zheng, Jiangbin
AU - Zhao, Danyang
AU - Ren, Jin Chang
PY - 2015
Y1 - 2015
N2 - Depth Image Based Rendering (DIBR), as a freeviewpoint synthesis technique, enables interactive selection of the view for watching. However, many rendering methods based on DIBR usually bring contour, crack and disocclusion artifacts. To address these problems, we propose effective methods to remove these artifacts. Firstly, before warping, the reference depth maps and color images are analyzed to find the regions causing contour artifacts. A combination of depth map edge detection and color consistency correction is applied to the analysis. By omitting warping the found regions, the synthesized views contain no contour artifacts and edges of foreground objects are well preserved. Secondly, cracks are filled using surrounding noncrack pixels rather than filtering for local consistency and smoothness. Thirdly, we apply a method based on texture extrapolation with depth information to inpaint the disocclusions. For two well-known sequences, 'Ballet' and Breakdancers', we obtain large Peak Signal to Noise Ratio (PSNR) gains in comparison to state-of-the-art techniques. In addition, the proposed method also obtains good results in Structural Similarity Index Measurement (SSIM) and visual quality.
AB - Depth Image Based Rendering (DIBR), as a freeviewpoint synthesis technique, enables interactive selection of the view for watching. However, many rendering methods based on DIBR usually bring contour, crack and disocclusion artifacts. To address these problems, we propose effective methods to remove these artifacts. Firstly, before warping, the reference depth maps and color images are analyzed to find the regions causing contour artifacts. A combination of depth map edge detection and color consistency correction is applied to the analysis. By omitting warping the found regions, the synthesized views contain no contour artifacts and edges of foreground objects are well preserved. Secondly, cracks are filled using surrounding noncrack pixels rather than filtering for local consistency and smoothness. Thirdly, we apply a method based on texture extrapolation with depth information to inpaint the disocclusions. For two well-known sequences, 'Ballet' and Breakdancers', we obtain large Peak Signal to Noise Ratio (PSNR) gains in comparison to state-of-the-art techniques. In addition, the proposed method also obtains good results in Structural Similarity Index Measurement (SSIM) and visual quality.
KW - Color consistency correction
KW - Contour artifacts
KW - Crack artifacts
KW - Depth Image Based Rendering (DIBR)
KW - Depth map edge detection
KW - Disocclusion artifacts
UR - http://www.scopus.com/inward/record.url?scp=84969765329&partnerID=8YFLogxK
U2 - 10.18293/DMS2015-014
DO - 10.18293/DMS2015-014
M3 - 会议稿件
AN - SCOPUS:84969765329
T3 - Proceedings - DMS 2015: 21st International Conference on Distributed Multimedia Systems
SP - 65
EP - 71
BT - Proceedings - DMS 2015
PB - Knowledge Systems Institute Graduate School
T2 - 21st International Conference on Distributed Multimedia Systems, DMS 2015
Y2 - 31 August 2015 through 2 September 2015
ER -