1. Adaptive Rood Pattern Search for Fast Block-Matching Motion Estimation Yao Nie, Student Member, IEEE, and Kai-Kuang Ma, Senior Member, IEEE IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 11, NO. 12, DECEMBER 2002 Abdul Arfan 20101204005
2. Outline (1) Background Full Search BMA Matching error surface 4 Typesof Fast BMA Pattern Adaptive pattern Prediction of target MV Fixed Pattern Zero Motion Prejudgment
3. Outline (2) The Complete Algorithm A checking bit-map Animation Experimental Result Implementation Other Approach Implementation Evaluation Evaluation Result Conclusion
4. Background Block-Matching algorithm (BMA) for motion estimation (ME) has been widely adopted by current video coding standard H.261, H.263, H.264, MPEG-1, MPEG-2, MPEG-4, and H.264 The most straightforward BMA is Full Search Exhaustively search for the best matching block within the search window Multimedia Class Project Presentation 4
5. Full Search BMA Very high computational complexity Makes the ME the main bottleneck in real-time video coding application Need faster BMA Algorithm Multimedia Class Project Presentation 5 [0]
8. 4 Types of Fast BMA: Fast BMA using a Fixed Set of Search Patterns Fast BMA based on Inter-Block Correlation Fast BMA using Hierarchical or Multi-resolution Search Framework Fast BMA using Sub-sampled Pixels on Matching-Error Computations Multimedia Class Project Presentation 8
9. Pattern Adaptive pattern Size can change depends on prediction of target MV Fixed Pattern Fixed size
10. Adaptive pattern For the initial search Check the best point among five points in the pattern Size of pattern can change based on predicted MV
11. Prediction of target MV (1) Spatial From neighboring MB Temporal From previous frame
16. Fixed Pattern For refined local search 2 patterns introduced The second gives similar PSNR, but require 40%-80% more checking points
17. Zero Motion Prejudgment If the SAD of macro block using motion vector 0,0 is less than T, then no searching is necessary. T value is 512
18. The complete algorithm (1) Step 1: Compute the matching error (SAD ) between the current block and the block at the same location in the reference frame (i.e., the center of the current search window).
19. The complete algorithm (2) Step 2: Align the center of ARP with the center point of the search window and check its four search points plus the position of the predicted MV to find out the current MME (minimal matching error) point.
20. The complete algorithm (3) Step 3: Set the center point of the unit-size rood pattern (URP) at the MME point found in the previous step and check its points. If the new MME point is not incurred at the center of the current URP, repeat this step; otherwise, the MV is found, corresponding to the MME point identified in this step.
21. A checking bit-map A checking bit-map (one bit for denoting the status of each macro block) has been employed to record whether a search point under checking has already been examined before, so that duplicated checking computation can be avoided.
23. Experimental Result (1) Simulations based on the encoding platform, MoMuSys FCD version 2.0.2, under MPEG-4 test conditions Compared with FS, ARPS greatly improves the search speed with computational gain in the range of 94 to 447. Meanwhile, ARPS maintains similar PSNR performance of FS in most sequences with less than 0.12 dB degradation When compared with DS, ARPS is constantly around 2 times faster with similar PSNR achieved.
28. Implementation Java Programming Language Code based on homework 1 of multimedia course Modified : A checking bit-map is changed to nested Hashtable The code runs much faster compared to full search, the resulting motion estimation video remains similar
29. Other Approach To compare the result Implements Full Search with search range ±7
32. Conclusion ARP adaptively exploits adjustable rood-shaped search pattern (which is powerful in tracking motion trend), together with the search point indicated by the predicted MV. Zero-motion prejudgment (ZMP) is incorporated into ARPS to further benefit small motion video sequence ARPS–ZMP improves average PSNR performance in large motion video sequences (e.g., 0.24 dB higher in Foreman and 0.39 dB higher in Coastguard).
