Presentation on theme: "Wen-Hsiao Peng Chun-Chi Chen"— Presentation transcript:
1 Wen-Hsiao Peng Chun-Chi Chen An Inter-Frame Prediction Technique Combining Template Matching Prediction and Block Motion Compensation for High Efﬁciency Video CodingCircuits and Systems for Video Technology, 2013 IEEE Transactions onWen-Hsiao PengChun-Chi Chen
2 Outline Introduction Background Bi-prediction Combining TMP and BMC Analysis LS and LMSExperiment ResultsConclusion
3 Introduction Inter prediction combines MVs from TMPBMCfor Overlapped Block Motion Compensation.Prediction performance of OBMC close to that of bi-prediction.without having to signal the template MV
4 Introduction TMP generally outperforms SKIP prediction. TMP is inferior to block-based motion compensation.Another MV is required to best complement the template MV.
5 IntroductionA key issue in video coders with motion-compensated prediction is how to trade off effectively betweenaccuracy of the motion field representationrequired overheadBased on HEVC version 6.0Achieve the bitrate reduction.
6 Outline Introduction Background Bi-prediction Combining TMP and BMC Template Matching PredictionBlock Motion CompensationSKIP and Merge-SKIPSignal ModelPrediction Error SurfaceBi-prediction Combining TMP and BMCAnalysis LS and LMSExperiment ResultsConclusion
7 Template Matching Prediction Obtains the MV at a current pixel by finding, in the reference frames, the best match for a template region composed of its surrounding reconstructed pixels.
8 Block Motion Compensation The frames are partitioned in blocks of pixels and each block is predicted from a block of equal size in the reference frame.
10 SKIP and Merge-SKIPSKIPH.264/AVCMerge-SKIPWeighted sum
11 Signal Model Tao et al  Zheng et al  .  B. Tao and M. T. Orchard, “A parametric solution for optimal overlapped block motion compensation,”IEEE Trans. on Image Processing, vol. 10, no. 3, pp. 341–350, Mar W. Zheng, Y. Shishikui, M. Naemura, Y. Kanatsugu, and S. Itoh,“Analysis of space-dependent characteristics of motion-compensated frame differences based on a statistical motion distribution model,”IEEE Trans. on Image Processing, vol. 11, no. 4, pp. 377–386, Apr
12 Signal Model Mean-sqaured prediction error Tao et al  .Tao et al Zheng et al  B. Tao and M. T. Orchard, “A parametric solution for optimal overlapped block motion compensation,”IEEE Trans. on Image Processing, vol. 10, no. 3, pp. 341–350, Mar W. Zheng, Y. Shishikui, M. Naemura, Y. Kanatsugu, and S. Itoh,“Analysis of space-dependent characteristics of motion-compensated frame differences based on a statistical motion distribution model,”IEEE Trans. on Image Processing, vol. 11, no. 4, pp. 377–386, Apr
13 Signal Model Block MV, vb , and block center, sc vb = v(sc).Template MV, vt , and template center, stvt = v(st)
15 Prediction Error Surface The error variance tends to be larger at the block boundaries and smaller around the center, which is understandable if we recall that vb approximates v(sc), the true motion associated with the block center.TMP better than BMC => if the intensity and motion fields are less random or have a high spatial correlation
16 Prediction Performance Comparsion Encoding 50 framesBasketballDrill has complex motionJohnny is of video-conferencing type and has less detail
17 Outline Introduction Background Bi-prediction Combining TMP and BMC Overlapped Block Motion CompensationLeast Square SolutionLeast Mean-Square SolutionAnalysis LS and LMSExperiment ResultsConclusion
18 Bi-prediction Combining TMP and BMC Predictor is computed as a weighted average of two reference blocks.Template MV, vtBlock MV, vbTMP can better compensate for the movement of the top-left area of a prediction block.BMC is thus aimed at reducing further the prediction residual in the remaining area.
19 Overlapped Block Motion Compensation The weighting can be pixel adaptive..ω is indicating their likelihoodThe problem is to determine the OBMC weights so that the resulting predictor would produce a minimal residual.
20 Overlapped Block Motion Compensation How to minimize the prediction residual by a suitable choice of the block MV and OBMC weights..The approaches to solve the problemLeast Squares ApproachLeast Mean-Square Approach
21 Least Square SolutionRely on an iterative algorithm to solve for the optimal weights.Estimating Block MVs :.Adapting OBMC Weights :It’s convergence to a possibly local minimum is usually between 5 to 10 iterations.
22 Least Mean-Square Solution Introduce statistical signal models.Given that every block is to be predicted using OBMC based on two MVsdefaulting to the true MVMV sampling the motion field at some point sbdetermine a set of OBMC weightsTemplate MV approximating the pixel true motion can be obtained first at the template centroid.better compensate for top-left cornerThe optimal choice for the other MV is given by the true motion of a pixel in the bottom-right quadrant.
23 Least Mean-Square Solution Transform the problem of minimizing ξ into that of minimizing its expected value E[ξ]..Fixing sb determine the :Find the optimal sb that yields the global minimum :
24 Outline Introduction Background Bi-prediction Combining TMP and BMC Analysis LS and LMSExperiment ResultsConclusion
25 Analysis LS and LMSindicates the likelihood of vt being the true motion of a pixel at s relative to the other hypothesis vb.Template MV is not as reliable for compensating pixels in the upper-left area as predicted by the theoretical results.Tao’s modelZheng’s modelLS solution
26 Multiple reference frames Analysis LS and LMSSo, we would expect to drop to zero (or, equivalently, to increase to unity)without amendmentwith amendmentMultiple reference frames
33 Outline Introduction Background Bi-prediction Combining TMP and BMC Analysis LS and LMSExperiment ResultsConclusion
34 ConclusionWe proposed a bi-prediction scheme that combines BMC and TMP predictors through OBMC.TMP is inferior to BMC, but is, in general, superior to SKIP prediction.The data dependency complicates the pipeline design and hinders parallel processing.The proposed method restricted the use of TB-mode to 2Nx2N PUs only.For TMP to work properly, pixels in the template region must be reconstructed prior to the motion estimation and compensation of a current PU.