The GPDM extends the Gaussian Process Latent Variable Model (GP-LVM) to data with temporal dependencies. In all the examples below we train either single or multiple-style models using data from the subject 137 in the CMU Motion Capture Database. This is the same dataset used for the CRBM experiments in Section 3.5.4.
For each model and style we show three sequences: 1) a sequence generated from the same initialization as we used for the CRBM; 2) the best sequence, as determined by visual inspection, over ten different initializations spaced uniformly over the training data; and 3) reconstructing the training data using the latent representation.
We used the GPDM implementation provided by Neil Lawrence in his FGPLVM toolbox. We train a single GPDM per-style. Hyperparameters of the dynamics GP are set by hand.
The player will show in this paragraph unless you do not have flash player installed
We used the GPDM implementation provided by Jack Wang on his GPDM website. We train a single balanced GPDM per-style. Hyperparameters of the dynamics GP are optimized.
The player will show in this paragraph unless you do not have flash player installed
We used the GPDM implementation provided by Neil Lawrence in his FGPLVM toolbox. Here, a single GPDM is trained on the 10-style dataset. Hyperparameters of the dynamics GP are set by hand.
The player will show in this paragraph unless you do not have flash player installed
We used the GPDM implementation provided by Jack Wang on his GPDM website. Here, a single balanced GPDM is trained on the 10-style dataset. We train a single model per-style. Hyperparameters of the dynamics GP are optimized.
The player will show in this paragraph unless you do not have flash player installed