Structured Forests for Fast Edge Detection

Structured Forests for Fast Edge DetectionPiotr Dollr and Larry Zitnick

1what defines an edge?BrightnessColorTextureParallelismContinuitySymmetry

Let the data speak.21. Accuracy2. Speed3I. data driven edge detection4edge detection as classificationSupervised Learning of Edges and Object BoundariesCVPR 2006, Piotr Dollr, Zhuowen Tu, Serge Belongiepositives

Hard!{ 0, 1 }

5edge have structure

6sketch tokensSketch Tokens, CVPR 2013. Joseph Lim, C. Zitnick, and P. Dollr

7random forests

8upgrading the output space{ 0, 1 }{ }

dimensionality222561519II. structured edge learning10structured forestsStructured Class-Labels in Random Forests for Semantic Image Labelling, ICCV 2011,P. Kontschieder, S. Rota Bul, H. Bischof, M. Pelillo

11tree training12node traininghigh entropy split low entropy split 13how to train?bad split

good split ??

14

??clusterminimize entropy

15III. structured edge detection16structured forests

17sliding window detector

18sliding window detectorpixel output structured output

Comparison to storing only single pixel predictionNoisy, no coherent structureWould need far more trees!This is the secret to our speed / accuracy.

19multiscale detection

20

x2 x

1 xmultiscale detection++21IV. results22

SS=single-scale MS=multi-scale T=# trees6HzSS T=1SS T=4MS T=4ODS = 0.72ODS = 0.74ODS = 0.7360Hz30HzgPb ODS=.73 gPb ODS=.73 FPS 1/240 Hz23

24

thanks!source code available online