Motion Map: Image-based Retrieval and Segmentation of Motion Data EG SCA ’ 04 學生 : 林家如
Outline Introduction Framework Results Conclusions Future Works
Introduction Semantic-based retrieval lacks the capability of accurately clipping the proper segment of the data. Provide GUI for retrieving motion data. Using Self-organizing map (SOM).
Introduction Only need to specify starting and ending postures.
Motion Map Constructing a graphical user interface for motion data retrieval.
SOM Self-organizing feature map network. A type of unsupervised learning. Usually 1D or 2D. A mapping that preserves neighborhood relations. Often used in information visualization.
SOM For each sample posture, an input vector is defined as model vector, m i,j
SOM model vector Learning-rate: The width of kernel:
Clustering Divides regions by detecting borders The average difference against 4 neighbors Create vertical border if Labeling
Posture Icons From the node that is nearest to the center of each clustered region.
Trajectory Each motion can be represented as a trajectory. The walking motion:
Virtual Node Increase the resolution with small computational cost. Can be preprocessed for great detail with the cost of storage.
Retrieval
Results
Conclusions Contributions: Automatically Easily Retrieve Display motion as a trajectory Defects: Can ’ t distinguish different performers Can ’ t reflect the dynamical feature
Future Works Analyzing minute difference. Zooming in the motion trajectories. Interactive data editing. Motion blending by drawing an interpolation path on the map.