Dlib -> PlugFrame Converter 2.0

2018.4

Nov 29

The Problems:

[1] Localized clip normalization doesn’t produce consistent results
[2] Identifying range of motion for normalization requires consistent point data
[3] Mass averaging creates soft calibration data

When I wrote the initial dlib converter in the spring of 2018 (with some initial help from Antonio and Jesse), it was done quickly in order to get it up and running in time for demos. At that time, we had a single Mindtwin - Obama. So, the converter was written to make PlugFrames from a set of dlib points in order to work with the bulk of clips of Reggie, our Obama impersonator. Hacks were employed in order to translate data to the puppet in the form of PlugFrames.

A couple months later I revised the converter slightly to use calibration data to normalize dlib data on a per-sequence basis. This provided much tighter results, but they were ultimately inconsistent since a longer clip containing many frames would provide a better range of motion for normalization as opposed to a short clip. In fact, a short clip may never hit the bounds of normalization that are required.