Optimal Motion Vector Search Algorithm 6 th Team











- Slides: 11

Optimal Motion Vector Search Algorithm 6 th Team 20032026 Kim, Hyun-Seok 20032072 Jang, Sun-Yean 20032077 Jung, Yu-Chul Digital Media Lab Concrete Mathematics KIM, HYUNSEOK / JANG, SUNYEAN / JUNG , YUCHUL

☆ Overview ☆ - Terminology - Block Matching - Motion Vector Search Algorithms - Considering Points - Our Suggestion - Implementation Outline - References Digital Media Lab Concrete Mathematics KIM, HYUNSEOK / JANG, SUNYEAN / JUNG , YUCHUL

☆ Terminology ☆ • Reference Frame : Frame in past (or future) used to predict in current frame • Current Frame: Frame which is being analyzed to derive motion vectors • Motion vector : The displacement of the closest matching block in reference frame for a block in current frame • Motion Estimator : Process of determining the values of motion vectors for each frame Digital Media Lab Concrete Mathematics KIM, HYUNSEOK / JANG, SUNYEAN / JUNG , YUCHUL

☆ Block Matching ☆ -to find the “best” block from an earlier frame to construct an area of the current frame -Goal is to find a vector where MAD(Mean Absolute Difference) is minimum. Digital Media Lab Concrete Mathematics KIM, HYUNSEOK / JANG, SUNYEAN / JUNG , YUCHUL

☆ Previous Approaches ☆ I. Full Search II. III. - Every possible block in the previous frame is examined - Of all the blocks examined, the lowest MAE produced is chosen, and the motion vector from that block’s position to the current block’s position is generated. - Problems : The most precise matching, but the most demanding in terms of computational complexity. (2 w+1)^2 times IV. Digital Media Lab Concrete Mathematics KIM, HYUNSEOK / JANG, SUNYEAN / JUNG , YUCHUL

☆ Previous Approaches(cont’) ☆ 1. II. Conjugate Direction Search 2. 3. - based on the assumption that the energy of the prediction error is monotonically decreasing towards the optimum motion vector in the search range. 1) first, search along the horizontal row of blocks in the previous frame The MAD is computed between each of these blocks 2) Then, extent the search in the vertical direction, searching the column of blocks in the previous frame which have the same xcoordinate as the best matching block founded in step 1) 4. 5. 6. 7. 8. 9. Comparing with the full search, complexity is reduced noticeably 3+2 w Digital Media Lab Concrete Mathematics KIM, HYUNSEOK / JANG, SUNYEAN / JUNG , YUCHUL

☆ Previous Approaches (cont’) ☆ 1. III. Modified Logarithmic Search 2. 3. - efficient and fast 2+7 log(w) - unable to search all of the locations at the boundaries of the search window, thus, it doesn’t always result in the optimum notion vector within the search window. However, its performance is very good for small displacements. Digital Media Lab Concrete Mathematics KIM, HYUNSEOK / JANG, SUNYEAN / JUNG , YUCHUL

☆ Considering Points ☆ - Full search is simple and correct, but computational burden. - Other approaches are apt to get trapped in local minima, resulting in a significant loss in estimation accuracy, and compression performance in video coding, as compared to the Full search ☆ What is needed? - Novel motion vector prediction technique - A highly localized search pattern - A computational constraint explicitly incorporated into the cost measure Digital Media Lab Concrete Mathematics KIM, HYUNSEOK / JANG, SUNYEAN / JUNG , YUCHUL

☆ Our Suggestion ☆ Concepts 1. Employ a representative value based on bit information : To maximize the correctness in potential motion changes 2. Use memory hash table : To reduce computational time 3. Use Nearest Neighbor hood Algorithm : To reduce the possibility of getting trapped in local minima Digital Media Lab Concrete Mathematics KIM, HYUNSEOK / JANG, SUNYEAN / JUNG , YUCHUL


☆ References ☆ 1. Correlation Based Search Algorithms for Motion Estimation Mohamed Alkanhal, Deepak Turgaga and Tsuhan Chen – E/CE of CMU, USA (Picture Coding Symposium Portland, OR, April 21~23, 1999) 2. An Efficient Computation-Constrained Block-Based Motion Estimation Algorithm for Low Bit Rate Video Coding Michael Gallant and Faouzi Kossentini – E/CE of UBC, Canada 3. Motion Vector Refinement for High-Performance Transcoding Jeongnam Youn, Ming-Ting Sun, Fellow, IEEE, and Chia-Wen Lin IEEE Transaction on Multimedia, Vol. 1, No. 1, March 1999 4. Computation constrained fast-search motion estimation algorithm for TMN 7. In Q 15 -A-45, ITU-T Q 15/SG 16, Portland, Oregon, June 1997 5. http: //www. dcs. warwick. ac. uk/research/mcg/bmmc/index. html Digital Media Lab Concrete Mathematics KIM, HYUNSEOK / JANG, SUNYEAN / JUNG , YUCHUL