Browsing by Author "Nishigaki, Morimichi"

Now showing 1 - 2 of 2
  • Thumbnail Image
    Item
    Ego-Motion Estimation using Fewer Image Feature Points
    (2004) Nishigaki, Morimichi; ISR
    Environmental recognition using images is a worthwhile research subject. 3D information is valuable information to recognize surrounding, because currently it is difficult for machines to understand 3D structure in the scene with only one image, even though a human can understand the scene structure from a picture. 3D information can be obtained from stereo and motion disparities. Since the stereo camera is assumed to be calibrated, the 3D shape can be calculated from disparities. On the other ahnd, since the camera motion, so called ego-motion, is not known, even though motion disparities is obtained,3D shape cannot be calculated until ego-motion is estimated. In addition, the 3D shape recovered from motion disparities is determied up to a scale. Generally, the accuracy of 3D reconstruction by stereo camera depends on the baseline. A stereo camera usually cannot take long baseline, while camera motion can produce long baseline. Since each stereo and motion has information on 3D shape, the combination of stereo and motion disparities could complementally produce better 3D reconstruction than only one of them. Researches in this report were taken placed toward the 3D reconstruction using stereo and motion information.
  • Thumbnail Image
    Item
    THE IMAGE TORQUE OPERATOR FOR MID-LEVEL VISION: THEORY AND EXPERIMENT
    (2012) Nishigaki, Morimichi; Aloimonos, Yiannis; Fermuller, Cornelia; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)
    A problem central to visual scene understanding and computer vision is to extract semantically meaningful parts of images. A visual scene consists of objects, and the objects and parts of objects are delineated from their surrounding by closed contours. In this thesis a new bottom-up visual operator, called the Torque operator, which captures the concept of closed contours is introduced. Its computation is inspired by the mechanical definition of torque or moment of force, and applied to image edges. It takes as input edges and computes over regions of different size a measure of how well the edges are aligned to form a closed, convex contour. The torque operator is by definition scale independent, and can be seen as an operator of mid-level vision that captures the organizational concept of 'closure' and grouping mechanism of edges. In this thesis, fundamental properties of the torque measure are studied, and experiments are performed to demonstrate and verify that it can be made a useful tool for a variety of applications, including visual attention, segmentation, and boundary edge detection.