Functional Optimal Transport:
Mapping Estimation and Domain Adaptation for Functional data
1. Mapping sample functions from one domain to another
Source data:
Swiss-roll curves
Target data:
Wave curves
The geodesic of the Mapping
The resulting
push-forward
2. Toy examples
2.1 From one high variance mixture to a small variance mixture
Source & Target data
The geodesic of the Mapping
The probabilistic coupling
(Joint distribution)
The Lambda matrix
of the Mapping
2.2 From two mixtures to two mixtures
Source & Target data
The geodesic of the Mapping
The probabilistic coupling
(Joint distribution)
The Lambda matrix
of the Mapping
2.3 From one mixture to two mixtures
Source & Target data
The geodesic of the Mapping
The probabilistic coupling
(Joint distribution)
The Lambda matrix
of the Mapping
3. Using FOT to map robot-arm demonstrations from one domain to another
3.1 Source: "bins-Bread" (Roboturk dataset), Target: "Picking-left-arm" (MIME dataset)
Source: Roboturk bins-Bread
Exaggerated motion
Target: MIME Picking-left-arm
Gently picking
FOT Push forward
Exaggerated picking
3.2 Source: "pegs-RoundNut" (Roboturk dataset), Target: "Pouring-left-arm" (MIME dataset)
Source: Roboturk pegs-RoundNut
Relative slow motion
Target: MIME Pouring-left-arm
Pick and then pour
FOT Push forward
Conservative motion