Active Segmentation in 3D using Kinect Sensor

The combination of color image and depth map significantly improves the segmentation. The Kinect sensor with pan/tilt motorized movement captures both images and segments them separately by the Grab Cut method. The resulting contours are converted to polar coordinates. After the floor plane detection, corresponding "depth" and "color" contours are combined such that the importance of depth /color information is proportional to the distance from the floor. The segmentation is followed by the extraction of simple scale invariant features like color components and height/width ratio. Subsequently, features are used to train Normal Bayes Classifier. The algorithm was tested on a set of simple objects (mugs) on the table.