Computer Vision News - November 2016

6 Computer Vision News Research Research ( ) is the loss for each scale ∈ {1, . . . , 5} denotes the frozen “base CNN” parameters. { ( ) , = 1, . . . , 5} the corresponding weights of the m-th activation side output. The multiplier is used to handle the imbalance of the substantially greater number of the background compared to the contour pixels. More details on the equation and all of its parameters can be found in the paper. (III) Estimation of Contour Orientations: For predicting the contour orientations, a function consisting of 8 sub-networks, each associated with one orientation, takes as its input the 5 side-outputs of the “base CNN”. Specifically, the orientation map is obtained as: B k (x, y) denotes the response of the k-th orientation bin of the CNN at the pixels with coordinates (x, y) and T (·) is the transformation function which associates each bin with its central angle. The learned contour orientations for the computation of the Oriented Watershed Transform (OWT), further boost performance. After the network was trained from a single pass of a base CNN, COB obtains multiscale oriented contours. COB combines them to build Ultrametric Contour Maps (UCMs) at different scales and fuse them into a single hierarchical segmentation structure.