segmenting motion capture data into distinct behaviors
DESCRIPTION
Segmenting Motion Capture Data into Distinct Behaviors. Graphics Interface ‘ 04 Speaker: Alvin January 17, 2005. Outline. Introduction Related Work PCA PPCA GMM Results Conclusions. Introduction. Motion data are segmented at capture or by hand and are often small clips. - PowerPoint PPT PresentationTRANSCRIPT
![Page 1: Segmenting Motion Capture Data into Distinct Behaviors](https://reader036.vdocument.in/reader036/viewer/2022062500/56815996550346895dc6de25/html5/thumbnails/1.jpg)
Segmenting Motion Capture Data into Distinct Behaviors
Graphics Interface ‘04
Speaker: Alvin
January 17, 2005
![Page 2: Segmenting Motion Capture Data into Distinct Behaviors](https://reader036.vdocument.in/reader036/viewer/2022062500/56815996550346895dc6de25/html5/thumbnails/2.jpg)
Alvin/GAME Lab./CSIE/NDHU
Segmenting Motion Capture Data into Distinct Behaviors
2
Outline
• Introduction
• Related Work
• PCA
• PPCA
• GMM
• Results
• Conclusions
![Page 3: Segmenting Motion Capture Data into Distinct Behaviors](https://reader036.vdocument.in/reader036/viewer/2022062500/56815996550346895dc6de25/html5/thumbnails/3.jpg)
Alvin/GAME Lab./CSIE/NDHU
Segmenting Motion Capture Data into Distinct Behaviors
3
Introduction
• Motion data are segmented at capture or by hand and are often small clips.
• Longer shots contain natural transitions.
• Segment motion into high-level behaviors.
• Unsupervised Learning
• Focus on efficient techniques: PCA, PPCA and GMM.
![Page 4: Segmenting Motion Capture Data into Distinct Behaviors](https://reader036.vdocument.in/reader036/viewer/2022062500/56815996550346895dc6de25/html5/thumbnails/4.jpg)
Alvin/GAME Lab./CSIE/NDHU
Segmenting Motion Capture Data into Distinct Behaviors
4
Related Work
• Model-based Approach• Low-level
• Detect zero crossings of angular velocities.• Motion texton• State Machine or Motion Graph
• High-level• HMM• Clustering
![Page 5: Segmenting Motion Capture Data into Distinct Behaviors](https://reader036.vdocument.in/reader036/viewer/2022062500/56815996550346895dc6de25/html5/thumbnails/5.jpg)
Alvin/GAME Lab./CSIE/NDHU
Segmenting Motion Capture Data into Distinct Behaviors
5
Goal
• Input: Motion data (14 motions, each 8000 frames)• FPS=120• 14 Joints• Specify the rotation relative to the parent for all joints.• Rotations are specified by quaternions.
• Output: Motion Clips• Automatically• Distinct Behaviors• Longer
![Page 6: Segmenting Motion Capture Data into Distinct Behaviors](https://reader036.vdocument.in/reader036/viewer/2022062500/56815996550346895dc6de25/html5/thumbnails/6.jpg)
Alvin/GAME Lab./CSIE/NDHU
Segmenting Motion Capture Data into Distinct Behaviors
6
Center of motion:
Approximation:
SVD:
Dimension:
Projection Error:
Derivative:
PCA
n
1iix
n
1x
rir2i21i1i v...vvx 'x
TV U D
56
1rj
2j
n
1i
2
'x-x iie
determine toframes 240 use 0.9,E 56
1j
2j
r
1j
2j
r
60 l,e-ed l-iii
![Page 7: Segmenting Motion Capture Data into Distinct Behaviors](https://reader036.vdocument.in/reader036/viewer/2022062500/56815996550346895dc6de25/html5/thumbnails/7.jpg)
Alvin/GAME Lab./CSIE/NDHU
Segmenting Motion Capture Data into Distinct Behaviors
7
PCA Cut if di more than 3 standard deviations from the average
![Page 8: Segmenting Motion Capture Data into Distinct Behaviors](https://reader036.vdocument.in/reader036/viewer/2022062500/56815996550346895dc6de25/html5/thumbnails/8.jpg)
Alvin/GAME Lab./CSIE/NDHU
Segmenting Motion Capture Data into Distinct Behaviors
8
Probabilistic PCA
• Average square of discard singular values:
• Covariance Matrix:• Average Mahalanobis
Distance• T=150, K=T• K:=K+ , =10, Thr△ △
eshold R=15
56
1ri
2i
2
r-56
1
T2VV
1-n
1C
TK
1Kii
1-Ti )x-(xC)x-(x
T
1H
![Page 9: Segmenting Motion Capture Data into Distinct Behaviors](https://reader036.vdocument.in/reader036/viewer/2022062500/56815996550346895dc6de25/html5/thumbnails/9.jpg)
Alvin/GAME Lab./CSIE/NDHU
Segmenting Motion Capture Data into Distinct Behaviors
9
PPCA
![Page 10: Segmenting Motion Capture Data into Distinct Behaviors](https://reader036.vdocument.in/reader036/viewer/2022062500/56815996550346895dc6de25/html5/thumbnails/10.jpg)
Alvin/GAME Lab./CSIE/NDHU
Segmenting Motion Capture Data into Distinct Behaviors
10
Gaussian Mixture Model
• Pre-processing:• Use PCA to project onto lower dimensional sub
space. (Speed up EM)• Preserve 90% of the variance.• Each cluster is represented by a Gaussian Distri
bution.
• EM• Estimate mean, covariance matrix, prior
![Page 11: Segmenting Motion Capture Data into Distinct Behaviors](https://reader036.vdocument.in/reader036/viewer/2022062500/56815996550346895dc6de25/html5/thumbnails/11.jpg)
Alvin/GAME Lab./CSIE/NDHU
Segmenting Motion Capture Data into Distinct Behaviors
11
GMM
![Page 12: Segmenting Motion Capture Data into Distinct Behaviors](https://reader036.vdocument.in/reader036/viewer/2022062500/56815996550346895dc6de25/html5/thumbnails/12.jpg)
Alvin/GAME Lab./CSIE/NDHU
Segmenting Motion Capture Data into Distinct Behaviors
12
GMM Cut if frames xi and xi+1 belong to different clusters
![Page 13: Segmenting Motion Capture Data into Distinct Behaviors](https://reader036.vdocument.in/reader036/viewer/2022062500/56815996550346895dc6de25/html5/thumbnails/13.jpg)
Alvin/GAME Lab./CSIE/NDHU
Segmenting Motion Capture Data into Distinct Behaviors
13
Results
Error Matrix for PCA Error Matrix for PPCA
![Page 14: Segmenting Motion Capture Data into Distinct Behaviors](https://reader036.vdocument.in/reader036/viewer/2022062500/56815996550346895dc6de25/html5/thumbnails/14.jpg)
Alvin/GAME Lab./CSIE/NDHU
Segmenting Motion Capture Data into Distinct Behaviors
14
Results
![Page 15: Segmenting Motion Capture Data into Distinct Behaviors](https://reader036.vdocument.in/reader036/viewer/2022062500/56815996550346895dc6de25/html5/thumbnails/15.jpg)
Alvin/GAME Lab./CSIE/NDHU
Segmenting Motion Capture Data into Distinct Behaviors
15
Results
Precision: Reported correct cuts / The total number of reported cuts
Recall : Reported correct cuts / The total number of correct cuts
![Page 16: Segmenting Motion Capture Data into Distinct Behaviors](https://reader036.vdocument.in/reader036/viewer/2022062500/56815996550346895dc6de25/html5/thumbnails/16.jpg)
Alvin/GAME Lab./CSIE/NDHU
Segmenting Motion Capture Data into Distinct Behaviors
16
Evaluation Form• 論文簡報部份
• 完整性介紹 (4)• 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 (4)• 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:
The mahalanobis distance can be adopted to my classification of motions. Besides, maybe I can exploit the GMM technique to classify for comparison.
![Page 17: Segmenting Motion Capture Data into Distinct Behaviors](https://reader036.vdocument.in/reader036/viewer/2022062500/56815996550346895dc6de25/html5/thumbnails/17.jpg)
Alvin/GAME Lab./CSIE/NDHU
Segmenting Motion Capture Data into Distinct Behaviors
17
Conclusions
• Imperfect because observations’ opinions.
• Treat all weights of DOF equally.
• Each method require some parameters.
• PCA-based methods work well.
• ICA may achieve better cut detection.
• No segmentation will apply for all applications.
![Page 18: Segmenting Motion Capture Data into Distinct Behaviors](https://reader036.vdocument.in/reader036/viewer/2022062500/56815996550346895dc6de25/html5/thumbnails/18.jpg)
Alvin/GAME Lab./CSIE/NDHU
Segmenting Motion Capture Data into Distinct Behaviors
18
Mahalanobis Distance• Dt(x) = (x – mt)S-1
t(x – mt)'• Dt is the distance from t group
• St represents the within-group covariance matrix
• mt is the vector of the means of t group
• X is the vector of frame values at location x
• Superior to Euclidean distance because it takes distribution of the points (correlations) into account
• Useful to determine the ”similarity” from an unknown sample to known samples
• Classify observations into different groups
![Page 19: Segmenting Motion Capture Data into Distinct Behaviors](https://reader036.vdocument.in/reader036/viewer/2022062500/56815996550346895dc6de25/html5/thumbnails/19.jpg)
Alvin/GAME Lab./CSIE/NDHU
Segmenting Motion Capture Data into Distinct Behaviors
19
GMM by Using EM