Figure 2.

Method Overview. With multi-view images as input (a), we first detect the 2D joints (b) and human semantic segmentation (c) results with learning method in every view image. After triangulating 2D joints to 3D, we could get all 3D pose seeds (d) with all connection of semantic neighbor joints. Then we reduce 3D pose seeds number through pre-assembling process (e). At last, through pose-assembling optimization, the final 3D poses (g) could be obtained combined with SMPL models (f) fitting from pre-assembling pose seeds.