Recently　emerged　RGB-D　sensors　provide　great　promise　for　indoor　scene　understanding,　which　is　a　fundamental　and　challenging　problem　in　computer　vision.　We　present　a　discriminative　model　in　this　paper　to　semantically　label　indoor　scenes　from　RGB-D　images　Unlike　previous　work　which　only　labels　pre-determined　superpixels,　we　characterize　the　scenes　with　a　set　of　planes　and　compose　them　into　objects.　The　optimal　way　to　composition　and　corresponding　labels　are　inferred　simultaneously　using　a　greedy　algorithm.　Our　model　considers　unary　features　and　pairwise　and　co-occurrence　context,　as　well　as　latent　variables　that　account　for　multi-mode　distributions　of　each　object　category.　We　train　the　model　with　latent　structural　SVM　learning　framework.　Our　approach　achieves　state-of-the-art　performance　on　the　Cornell　RGB-D　indoor　scene　dataset　[1].