In　this　paper,　we　proposed　a　segmentation　approach　that　not　only　segment　an　interest　object　but　also　label　different　semantic　parts　of　the　object,　where　a　discriminative　model　is　presented　to　describe　an　object　in　real　world　images　as　multiply,　disparate　and　correlative　parts.　We　propose　a　multi-stage　segmentation　approach　to　make　inference　on　the　segments　of　an　object.　Then　we　train　it　under　the　latent　structural　SVM　learning　framework.　Then,　we　showed　that　our　method　boost　an　average　increase　of　about　5%　on　ETHZ　Shape　Classes　Dataset　and　4%　on　INRIA　horses　dataset.　Finally,　extensive　experiments　of　intricate　occlusion　on　INRIA　horses　dataset　show　that　the　approach　have　a　state　of　the　art　performance　in　the　condition　of　occlusion　and　deformation.　Copyright　©　2014　SCITEPRESS　-　Science　and　Technology　Publications.　All　rights　reserved.