The paper proposes a two-class classification model for image segmentation, using human-segmented images as positive examples and randomly matched human segmentations as negative examples. The preprocessing stage involves oversegmenting images into superpixels, which are then used to define various features derived from classical Gestalt cues, including contour, texture, brightness, and good continuation. Information-theoretic analysis is applied to evaluate the effectiveness of these grouping cues. A linear classifier is trained to combine these features. The segmentation problem is formulated as an optimization problem over the space of segmentations, and a simple algorithm is designed to randomly search for good segmentations. Experimental results on a wide range of images demonstrate the effectiveness of the classification model. The paper also discusses the power of the Gestalt cues and the evaluation of the classification model using precision-recall curves.The paper proposes a two-class classification model for image segmentation, using human-segmented images as positive examples and randomly matched human segmentations as negative examples. The preprocessing stage involves oversegmenting images into superpixels, which are then used to define various features derived from classical Gestalt cues, including contour, texture, brightness, and good continuation. Information-theoretic analysis is applied to evaluate the effectiveness of these grouping cues. A linear classifier is trained to combine these features. The segmentation problem is formulated as an optimization problem over the space of segmentations, and a simple algorithm is designed to randomly search for good segmentations. Experimental results on a wide range of images demonstrate the effectiveness of the classification model. The paper also discusses the power of the Gestalt cues and the evaluation of the classification model using precision-recall curves.