Structured Forests for ConBnuity Fast Edge Detection Symmetry - PowerPoint PPT Presentation

4/27/16 what defines an edge? Brightness Color Texture Parallelism Structured Forests for ConBnuity Fast Edge Detection Symmetry … Piotr Dollár and Larry Zitnick 1. A 1. Accur urac acy y 2. Speed 2. Speed I. data driven edge detec;on 1

4/27/16 edge detec;on as classifica;on edge have structure { 0, 1 } posiBves Supervised Learning of Edges and Object Boundaries CVPR 2006, Piotr Dollár, Zhuowen Tu, Serge Belongie sketch tokens random forests Sketch Tokens, CVPR 2013. Joseph Lim , C. Zitnick, and P. Dollár 2

4/27/16 upgrading the output space dimensionality 2 { 0, 1 } { … } 151 II. structured edge learning 2 256 structured forests tree training Structured Class-Labels in Random Forests for SemanBc Image Labelling, ICCV 2011, P. Kontschieder, S. Rota Bulò, H. Bischof, M. Pelillo 3

4/27/16 node training how to train? ? good split J ? bad split L low entropy split J high entropy split L minimize entropy cluster III. structured edge detec;on ? ? 4

4/27/16 structured forests sliding window detector sliding window detector mul;scale detec;on pixel output L structured output J 5

4/27/16 mul;scale detec;on + + ½ x 1 x IV. results 2 x SS T=1 ODS = 0.72 60Hz SS T=4 ODS = 0.73 30Hz MS T=4 ODS = 0.74 6Hz gPb ODS=.73 FPS ≈ 1/240 Hz SS=single-scale MS=mul8-scale T=# trees gPb ODS=.73 6

4/27/16 thanks! source code available online 7