Motion Texture: A Two-Level Statistical Model for Character Motion Synthesis SIGGRAPH ‘02 Speaker: Alvin Date: 3 August 2004
Alivn/GAME Lab./CSIE/NDHU 2 Outline Introduction Framework Result Conclusion Evaluation Form
Alivn/GAME Lab./CSIE/NDHU 3 Introduction Motion Texture – A two-level statistical model – Texton Local dynamics Represented by a linear dynamic system (LDS). – Distribution Global dynamics Modeled by a transition matrix – Counting how many times a texton is switched to another.
Alivn/GAME Lab./CSIE/NDHU 4 Two-level Statistical Model
Alivn/GAME Lab./CSIE/NDHU 5 Motion Texton State-Space model (LDS): X t – Hidden State Variable Y t – The Observation V t, W t – Independent Gaussian noises at time t.
Alivn/GAME Lab./CSIE/NDHU 6 Distribution Commonly used in HMMs
Alivn/GAME Lab./CSIE/NDHU 7 Introduction (cont.) Statistically similar to the original motion. Motion textures display a 1-D temporal distribution. User can synthesis and edit at both the texton level and the distribution level.
Alivn/GAME Lab./CSIE/NDHU 8 Framework Learning – E-step – M-step Synthesis – Texton Path Planning – Texton Synthesis By Sampling Noise With Constrained LDS
Alivn/GAME Lab./CSIE/NDHU 9 E-step Segment Labels as L = {l 1,l 2,…, l Ns } Segmentation points as H = {H 1,H 2,…,H Ns }
Alivn/GAME Lab./CSIE/NDHU 10 M-step
Alivn/GAME Lab./CSIE/NDHU 11 Learning Initialization - A greedy approach: Until the entire sequence is processed: – Use T min to fit LDS i – Label the subsequent frames to segment i until the fitting error is above a threshold. – Test all existing LDS’ to choose the best-fit LDS. – If no LDS fits well, introduce a new LDS.
Alivn/GAME Lab./CSIE/NDHU 12 Learning (cont.) The bigger the threshold, the longer the segments, and the fewer the number of textons. – Model selection methods: BIC MDL T min must be long enough to capture the local dynamics. – Approximately one second.
Alivn/GAME Lab./CSIE/NDHU 13 Texton Path Planning find a single best path,, which starts at and ends at. Two approaches: – Finding the Lowest Cost Path Dijkstra’s algorithm – Specifying the Path Length Dynamic Programming
Alivn/GAME Lab./CSIE/NDHU 14 Texton Synthesis By sampling noise – Inevitably depart from the original motion as time progresses. LDS learns only locally consistent motion patterns. The synthesis errors will accumulate as x t propagates. With constrained LDS – Setting the end constraints. – The in-between frames can be synthesized by solving a block-banded system of linear equations.
Alivn/GAME Lab./CSIE/NDHU 15 Result Environment – Intel P4 1.4G – 1G Memory Input – Capture 20 minutes of dance motion. (49800 frames) Result – It took about 4 hours to learn. – 246 textons are found. – The length of the texton ranges from 60 to 172 frames. – Synthesizing a texton only 25ms to 35ms. (Real-time)
Alivn/GAME Lab./CSIE/NDHU 16 Result (cont.)
Alivn/GAME Lab./CSIE/NDHU 17 Result (cont.)
Alivn/GAME Lab./CSIE/NDHU 18 Result (cont.)
Alivn/GAME Lab./CSIE/NDHU 19 Result (cont.)
Alivn/GAME Lab./CSIE/NDHU 20 Conclusion Best suited for repeated motions. Lack global variation when the data is limited. Did not incorporate any physical model into the synthesis algorithm. Capture the essential properties of the original motion.
Alivn/GAME Lab./CSIE/NDHU 21 Conclusion (cont.) The edited pose can not deviate from the original one too much. The additional constraint may contaminate the synthesized texton. Does not consider the interaction with environment. Initialization can be improved.
Alivn/GAME Lab./CSIE/NDHU 22 Evaluation Form 論文簡報部份 – 完整性介紹 (3) – 系統性介紹 (4) – 表達能力 (3) – 投影片製作 (3) 論文審閱部分 – 瞭解論文內容 (3) – 結果正確性與完整性 (4) – 原創性與重要性 (4) – 讀後啟發與應用: When we meet a problem that its input is highly repeating, we can use the statistical method to find the basic element.