BibTeX entry

@PHDTHESIS{201009Neill_Duncan_Francis_Campbell,
  AUTHOR={Neill Duncan Francis Campbell},
  TITLE={Automatic 3D Model Acquisition from Uncalibrated Images},
  SCHOOL={University of Cambridge},
  MONTH=Sep,
  YEAR=2010,
  URL={http://www.bmva.org/theses/2010/2010-campbell.pdf},
}

Abstract

The recovery of shape from images has received much attention as a core research topic in Computer Vision, however, these algorithms often require specialist equipment, expert knowledge or large numbers of images to obtain good results. These act as a barrier to the adoption of these technologies by other fields where we wish to allow systems to process images from the ‘real world’ without Computer Vision experts to operate them. The desire to implement a practical reconstruction system that is useful to and usable by the people who may benefit from these technologies motivates the contributions of this thesis. This work addresses all three of the stages required to take a sequence of images of an object and recover a 3D model in order to produce a system that maximises automation and minimises the demands placed on the user. To that end we present a practical implementation of an automatic method for recovering the positions and properties of the cameras used to take a series of images using a textured ground-plane. We then offer two contributions to simplify the task of segmenting an object observed in multiple images. The first, applicable to more simple scenes, automatically segments the object fixated upon by the camera. We achieve this by exploiting the rigid structure of the scene, to perform the segmentation in 3D across all the images simultaneously, and the consistent appearance of the object in an iterative method. For more complex scenes we move to our second algorithm that allows the user to select the required object in an interactive manner whilst minimising demands on their time. We combine the different appearance and spatial constraints to produce a clustering problem to group regions across images that allows the user to label many images at the same time. Finally we present an automatic reconstruction algorithm that improves the performance of existing state-of-the-art methods to allow accurate models to be obtained from smaller image sequences. This takes the form of a filtering process that rejects erroneous depth estimates by considering multiple depth hypotheses and identifying the true depth or an unknown state using a 2D Markov Random Field framework. We provide experimental validation for all the individual contributions, demonstrate the practical system working as a whole and conclude by discussing the merits of the system and avenues for future work.