Computer Vision News - June 2016

The model was trained based on 3 datasets:  Biomedical Image Library (BU-BIL:1-5), including 271 grayscale images coming from three fluorescence microscopy sets.  Weizmann, consisting of 100 grayscale images showing a variety of everyday objects.  Interactive Image Segmentation (IIS), with 151 RGB images showing a variety of everyday objects. “ The first study telling when to ‘Pull The Plug’ on humans or computers for image segmentation ” Feature Number of features Motivation for the feature set Segmentation Boundary 2 When algorithms fail, resulting segmentations often have boundaries characterized by an abnormally large proportion of highly-jagged edges Segmentation shape/Compactness 3 When algorithms fail, segmentations often are not compact Location of Segmentation in Image 2 When algorithms fail, resulting segmentation regions often lie closer to the edges of images Coverage of Segmentation on Image 2 When algorithms fail, resulting segmentations often cover abnormally large and small areas in the image Method In a broad sense, the Pull The Plug method is based on a regression model in order to capture the segmentation quality: from complete failures to nearly perfect (see Figure 1 in the previous page). The following settings were used to train the regression model:  Data - several segmentation masks were generated per training image: three variants of Hough Transform with Circles, Otsu Thresholding, adaptive thresholding and three binary masks based on the ground truth segmentations.  Labels - the Jaccard index used as indicator for the quality for each training instance segmentation.  Features - nine features derived from the binary segmentation mask aimed at capturing the segmentation quality. See the following table for more details about the observed features. 18 Computer Vision News Research