Computer Vision News - August 2016

As one can see, the area near the grass and the flowers contains many “false edges”. Ideally, we would prefer avoiding all edges caused by the grass and leaves. In fact, we aim at having only the edges between the houses, between the tree and the sky, between the grass and the houses, etc. In other words we would like the edges to ‘segment’ the image into semantically meaningful regions . With that aim in mind, we will now evaluate a much more meaningful and accurate image edge detection method, using the three techniques mentioned above. A. Structured Forests for Fast Edge Detection Computer Vision News Trick 13 This approach uses decision trees to robustly predict local edge masks in a structured learning framework (for more info see "Fast Edge Detection Using Structured Forests" ArXiv 2014 ). In order to use the “Structured Forests for Fast Edge Detection,” one first needs to train the model on some annotated ground truth dataset (if such exists). In case you don’t have one, the toolbox also comes with a pre- trained model which was trained on the BSDS500 dataset . The function to train the model (or to load the existing one) is named edgesTrain . Among the important parameter for this function are: nPos and nNeg used to set the number of negative and positive patches per tree respectively, which control the size of the tree in the models. Other parameters include ‘parfor’ which determines whether to parallelize the training process. Additional parameters are the number of trees in the model, the number of splits in each tree, and so on (see documentation for more details). With the following Matlab code snippet, invoke the edgesTrain and load the existing model (that comes with the toolbox), as specified by the modelDir parameter: Dear reader, How do you like Computer Vision News? Did you enjoy reading it? Give us feedback here: It will take you only 2 minutes to fill and it will help us give the computer vision community the great magazine it deserves! Give us feedback, please (click here) FEEDBACK Trick