Ideas

• visually detect contact (using proximity, shade, finger appearance (whitening?)

• as in SMPLify find 3D mesh for exisiting model and infer joints (a bit done in nyu-dataset where depth maps are generated from model to fit target)

• use hand parts (possibility to create 3D synthetic dataset)

• unsupervised approach for grasp classification

• consider that optimal grasp is used

• visually infer independent contact regions (see definition in grasps)