Weakly Supervised Body Part Parsing with Pose based Part Priors

The qualitative results on Pascal-Person-Part. No sample selection is conducted prior to the visualization, i.e. the displayed order strickly follows the validation index in Pascal-Person-Part. Input images are compressed due to the space limit.


Image

Target

Weakly Baseline Part Prior Supervised (43.91%)

Ours Weakly Base (60.35%)

Ours Weakly Iter (62.05%)

Fully Supervised (63.64%)

Ours Semi (68.88%)