Dense Descriptors Towards Fine Manipulation

:page_facing_up: abridged report

Directed Study - Fall 2021 - Professor Chad Jenkins

To train dense descriptors for fine object manipulation, we construct a more granular model and explore losses that encouraged the model to encode a sense of parts and poses.

The left image shows an auxiliary loss introduced to attempt to encourage coherent category level consistency of dense descriptors. To the right the image shows the resultant 3 dimensional dense descriptors.