Computer Vision News - November 2016

Computer Vision News Computer Vision News Research 5 Research the contour orientation is coupled with unique fast hierarchies form multiscale oriented contours to provide fast and accurate results obviating the need for the global and quite slow normalized-cut phase. Forth, experiments on BSDS, PASCAL Context, PASCAL Segmentation, and MS-COCO, demonstrate that COB provides state-of-the-art contours, region hierarchies, and object proposals in all datasets. Method: The COB consists of three main building blocks: (I) A “base CNN”, (II) Deep contour detectors, and (III) Estimation of contour orientations. A detailed description of each of those components follows. The figure below illustrates the components and their interactions. (I) A “base CNN”: the main building block of the COB is a pre-trained CNN. It could be either the 50-layer ResNet or the VGGNet. The given pre-trained network is adapted by removing the fully connected layers, and the batch normalization layers. The resulting pared-down network consists mainly of convolutional layers with ReLU activation function. This “base CNN” network is divided into 5 stages (activations), named “side outputs”, which produce scale-specific contour images. The “base CNN” consisting of 5 side-output stages is demarcated by a dashed red line in the figures. (II) Deep contour detectors: For the contour detectors, side-outputs 1-4 were combined to form while side-outputs 2-5 were combined to form COB trains the side-output detectors by separately supervising the output of the last layer of each, while freezing the pre-trained weights of the “base CNN”. The Deep contour detectors training is constructed using the following loss function: