Supplementary: AligningSilhouetteTopologyfor Self-Adaptive3DHumanPoseRecovery

It decodes a plausible pose when sampled inU[ 1,1]32 (green region) while sampling outside this bound may lead to implausible poses(redregion). Here, the Eq. 2 denotes generative adversarial loss on decoderΨ and the Eq. 3 specifies the loss ontheposediscriminatorDisc. Yellow ellipses highlight the region of articulation errors. We obtain pre-adaptation results of the source trained networks via direct inference on the shifted target (column 2 and 4 in panel A and B of Fig 3). We show these in order to qualitatively compare the improvement.