@InProceedings{Jenkins2003,
  author    = {Jenkins, O. C. and Mataric, M. J.},
  title     = {Automated Derivation of Behavior Vocabularies for Autonomous Humanoid Motion},
  booktitle = {Proceedings of the International Joint Conference on Autonomous Agents and Multiagent Systems},
  year      = {2003},
  pages     = {225--232},
  abstract  = {In this paper we address the problem of automatically deriving vocabularies of motion modules from human motion
data, taking advantage of the underlying spatio-temporal
structure in motion. We approach this problem with a
data-driven methodology for modularizing a motion stream
(or time-series of human motion) into a vocabulary of parameterized primitive motion modules and a set of metalevel behaviors characterizing extended combinations of the
primitives. Central to this methodology is the discovery of
spatio-temporal structure in a motion stream. We estimate
this structure by extending an existing nonlinear dimension
reduction technique, Isomap, to handle motion data with
spatial and temporal dependencies. The motion vocabularies derived by our methodology provide a substrate of autonomous behavior and can be used in a variety of applications. We demonstrate the utility of derived vocabularies for
the application of synthesizing new humanoid motion that
is structurally similar to the original demonstrated motion.},
  acmid     = {860612},
  doi       = {10.1145/860575.860612},
  groups    = {Lit Review 2013-09},
  isbn      = {1-58113-683-8},
  keywords  = {autonomous humanoid agents, humanoid robotics, kinematic motion segmentation, motion primitives, motion vocabularies, spectral dimension reduction},
  location  = {Melbourne, Australia},
  numpages  = {8},
  review    = {Want to auto-label human movement data. Uses spatial isomap.

Spatial isomap
1) bin spatial points together ("determine a local neighbourhood of nearby points") via k-nn
(don't quite follow the details here)
- basically, want to create a distance matrix D between data points and some neighbouring point
- this allows the isomap to be performed on this feature space similarity matrix D instead of some input space covariance matrix C (as is the case for PCA)
 - similar idea to Kernal PCA, where we only need the dot product to compute our similarity, as oppose to converting the entire input data space
 - they add temporal information by taking the Euclidean distance and warping the D matrix based on temporal neighbours
 - they like it because PCA only rotates the data, whereas isomap seems to do it more intelligently

After isomap warping, the segmentation is done by greedy algorithm:
1. current segment is the first frame
2. calculate distance between centroid from the current segment boundary (so t-1 frame) to every subsequent frame
3. pick first local maximum in centroid distance function

(so not online I guess) - results not reported



Jenkins and Matari\'{c} \cite{Jenkins2002, Jenkins2003} segments by performed by calculating the distance from the first frame's Cartesian centroid of a given DOF to all the subseqent frames, and segmenting on the first local maximum. Labeling and learning is done by performing dimensionality reduction using spatial Isomap by binning spatial data points together via \emph{k}-NN, computing the shortest distance between all point pairs using Dijkstra's algorithm, then preforming PCA. Although Isomap and PCA have the same general steps, Isomap is performed in the feature space, which is a higher dimensional space when compared to the input space, allowing for non-linearity in the input space. PCA works optimally if the input data is linear, or near linear. making Isomap a more flexible algorithm. Temporal information is accounted for by adding temporal-based weights to the distance metric. Similar activities are combined together by sweep-and-prune clustering. The focus of these paper is movement learning, and not segmentation, so no segmentation accuracy was reported.},
  timestamp = {2013.09.30},
}

{"_id":"YBihsxjBoH7a4HE4Y","bibbaseid":"jenkins-mataric-automatedderivationofbehaviorvocabulariesforautonomoushumanoidmotion-2003","downloads":0,"creationDate":"2017-09-14T16:34:36.577Z","title":"Automated Derivation of Behavior Vocabularies for Autonomous Humanoid Motion","author_short":["Jenkins, O. C.","Mataric, M. J."],"year":2003,"bibtype":"inproceedings","biburl":"https://raw.githubusercontent.com/jfslin/jfslin.github.io/master/jf2lin.bib","bibdata":{"bibtype":"inproceedings","type":"inproceedings","author":[{"propositions":[],"lastnames":["Jenkins"],"firstnames":["O.","C."],"suffixes":[]},{"propositions":[],"lastnames":["Mataric"],"firstnames":["M.","J."],"suffixes":[]}],"title":"Automated Derivation of Behavior Vocabularies for Autonomous Humanoid Motion","booktitle":"Proceedings of the International Joint Conference on Autonomous Agents and Multiagent Systems","year":"2003","pages":"225--232","abstract":"In this paper we address the problem of automatically deriving vocabularies of motion modules from human motion data, taking advantage of the underlying spatio-temporal structure in motion. We approach this problem with a data-driven methodology for modularizing a motion stream (or time-series of human motion) into a vocabulary of parameterized primitive motion modules and a set of metalevel behaviors characterizing extended combinations of the primitives. Central to this methodology is the discovery of spatio-temporal structure in a motion stream. We estimate this structure by extending an existing nonlinear dimension reduction technique, Isomap, to handle motion data with spatial and temporal dependencies. The motion vocabularies derived by our methodology provide a substrate of autonomous behavior and can be used in a variety of applications. We demonstrate the utility of derived vocabularies for the application of synthesizing new humanoid motion that is structurally similar to the original demonstrated motion.","acmid":"860612","doi":"10.1145/860575.860612","groups":"Lit Review 2013-09","isbn":"1-58113-683-8","keywords":"autonomous humanoid agents, humanoid robotics, kinematic motion segmentation, motion primitives, motion vocabularies, spectral dimension reduction","location":"Melbourne, Australia","numpages":"8","review":"Want to auto-label human movement data. Uses spatial isomap. Spatial isomap 1) bin spatial points together (\"determine a local neighbourhood of nearby points\") via k-nn (don't quite follow the details here) - basically, want to create a distance matrix D between data points and some neighbouring point - this allows the isomap to be performed on this feature space similarity matrix D instead of some input space covariance matrix C (as is the case for PCA) - similar idea to Kernal PCA, where we only need the dot product to compute our similarity, as oppose to converting the entire input data space - they add temporal information by taking the Euclidean distance and warping the D matrix based on temporal neighbours - they like it because PCA only rotates the data, whereas isomap seems to do it more intelligently After isomap warping, the segmentation is done by greedy algorithm: 1. current segment is the first frame 2. calculate distance between centroid from the current segment boundary (so t-1 frame) to every subsequent frame 3. pick first local maximum in centroid distance function (so not online I guess) - results not reported Jenkins and Matarić i̧teJenkins2002, Jenkins2003 segments by performed by calculating the distance from the first frame's Cartesian centroid of a given DOF to all the subseqent frames, and segmenting on the first local maximum. Labeling and learning is done by performing dimensionality reduction using spatial Isomap by binning spatial data points together via \\emphk-NN, computing the shortest distance between all point pairs using Dijkstra's algorithm, then preforming PCA. Although Isomap and PCA have the same general steps, Isomap is performed in the feature space, which is a higher dimensional space when compared to the input space, allowing for non-linearity in the input space. PCA works optimally if the input data is linear, or near linear. making Isomap a more flexible algorithm. Temporal information is accounted for by adding temporal-based weights to the distance metric. Similar activities are combined together by sweep-and-prune clustering. The focus of these paper is movement learning, and not segmentation, so no segmentation accuracy was reported.","timestamp":"2013.09.30","bibtex":"@InProceedings{Jenkins2003,\n author = {Jenkins, O. C. and Mataric, M. J.},\n title = {Automated Derivation of Behavior Vocabularies for Autonomous Humanoid Motion},\n booktitle = {Proceedings of the International Joint Conference on Autonomous Agents and Multiagent Systems},\n year = {2003},\n pages = {225--232},\n abstract = {In this paper we address the problem of automatically deriving vocabularies of motion modules from human motion\ndata, taking advantage of the underlying spatio-temporal\nstructure in motion. We approach this problem with a\ndata-driven methodology for modularizing a motion stream\n(or time-series of human motion) into a vocabulary of parameterized primitive motion modules and a set of metalevel behaviors characterizing extended combinations of the\nprimitives. Central to this methodology is the discovery of\nspatio-temporal structure in a motion stream. We estimate\nthis structure by extending an existing nonlinear dimension\nreduction technique, Isomap, to handle motion data with\nspatial and temporal dependencies. The motion vocabularies derived by our methodology provide a substrate of autonomous behavior and can be used in a variety of applications. We demonstrate the utility of derived vocabularies for\nthe application of synthesizing new humanoid motion that\nis structurally similar to the original demonstrated motion.},\n acmid = {860612},\n doi = {10.1145/860575.860612},\n groups = {Lit Review 2013-09},\n isbn = {1-58113-683-8},\n keywords = {autonomous humanoid agents, humanoid robotics, kinematic motion segmentation, motion primitives, motion vocabularies, spectral dimension reduction},\n location = {Melbourne, Australia},\n numpages = {8},\n review = {Want to auto-label human movement data. Uses spatial isomap.\n\nSpatial isomap\n1) bin spatial points together (\"determine a local neighbourhood of nearby points\") via k-nn\n(don't quite follow the details here)\n- basically, want to create a distance matrix D between data points and some neighbouring point\n- this allows the isomap to be performed on this feature space similarity matrix D instead of some input space covariance matrix C (as is the case for PCA)\n - similar idea to Kernal PCA, where we only need the dot product to compute our similarity, as oppose to converting the entire input data space\n - they add temporal information by taking the Euclidean distance and warping the D matrix based on temporal neighbours\n - they like it because PCA only rotates the data, whereas isomap seems to do it more intelligently\n\nAfter isomap warping, the segmentation is done by greedy algorithm:\n1. current segment is the first frame\n2. calculate distance between centroid from the current segment boundary (so t-1 frame) to every subsequent frame\n3. pick first local maximum in centroid distance function\n\n(so not online I guess) - results not reported\n\n\n\nJenkins and Matari\\'{c} \\cite{Jenkins2002, Jenkins2003} segments by performed by calculating the distance from the first frame's Cartesian centroid of a given DOF to all the subseqent frames, and segmenting on the first local maximum. Labeling and learning is done by performing dimensionality reduction using spatial Isomap by binning spatial data points together via \\emph{k}-NN, computing the shortest distance between all point pairs using Dijkstra's algorithm, then preforming PCA. Although Isomap and PCA have the same general steps, Isomap is performed in the feature space, which is a higher dimensional space when compared to the input space, allowing for non-linearity in the input space. PCA works optimally if the input data is linear, or near linear. making Isomap a more flexible algorithm. Temporal information is accounted for by adding temporal-based weights to the distance metric. Similar activities are combined together by sweep-and-prune clustering. The focus of these paper is movement learning, and not segmentation, so no segmentation accuracy was reported.},\n timestamp = {2013.09.30},\n}\n\n","author_short":["Jenkins, O. C.","Mataric, M. J."],"key":"Jenkins2003","id":"Jenkins2003","bibbaseid":"jenkins-mataric-automatedderivationofbehaviorvocabulariesforautonomoushumanoidmotion-2003","role":"author","urls":{},"keyword":["autonomous humanoid agents","humanoid robotics","kinematic motion segmentation","motion primitives","motion vocabularies","spectral dimension reduction"],"downloads":0},"search_terms":["automated","derivation","behavior","vocabularies","autonomous","humanoid","motion","jenkins","mataric"],"keywords":["autonomous humanoid agents","humanoid robotics","kinematic motion segmentation","motion primitives","motion vocabularies","spectral dimension reduction"],"authorIDs":[],"dataSources":["iCsmKnycRmHPxmhBd"]}