MSc thesis projects in Computational Vision at CB, CSC:

We perform basic research on image features in terms of receptive fields and their application to image matching and view-based object recognition. With close connections our research, we offer the following MSc thesis topics:

Local image descriptors for image matching and object recognition

A common approach to image matching and object recognition consists of detecting interest points and computing image descriptors at these points. Currently, the SIFT and SURF descriptors constitute the most frequently used image descriptors with a large number successful applications in computer vision and related fields. From local image measurements, it is however possible to compute other types of image descriptors that may lead to better performance for image matching and/or object recognition.

The subject of this MSc thesis project is to investigate, compare and if possible extend two recently developed approaches for computing local image descriptors from local image information, one based on Gaussian derivatives in terms of the local N-jet and one approach inspired by linear and non-linear receptive fields in biological vision. The formulation of this project is near basic research issues and opens up the possibility of comparing approaches in computer vision to computational approaches inspired by biological vision.

Contact: Tony Lindeberg (tony@csc.kth.se)

Efficient methods for matching collections of local image features

Given a set of local interest points with associated image descriptors that computed from an image and a set of image descriptors that have been pre-computed from the images in an image database, the object recognition problems can be expressed in terms of the problem of finding sets of image descriptors in the given image that match sets of image descriptors in the image database. The most basic approach to this problem consists of matching each image feature in the target image to all image features in the database based on nearest-neighbour matching. This approach may be sufficient, for example, when matching image features between the two images in a stereo pair or a small number of multiple views. For larger datasets, however, exhaustive nearest-neighbour search may be too time demanding, and more efficient search mechanisms are needed. When recognizing objects, groups of local image features do furthermore need to be collected to accumulate evidence for a particular object.

The subject of this MSc thesis project is to investigate, compare and if possible extend more efficient approaches to this problem. One possible candidate to consider is the tree-based approach by Nister and Stewenius where local image features are clustered into a vocabulary of visual words from which compact keys are computed and used for gathering evidence for different interpretations. A primary goal would then be to adapt and extend this approach to the generalized framework for interest point detection we are working with in our group and investigate the performance on benchmark datasets.

Contact: Tony Lindeberg (tony@csc.kth.se) or Oskar Linde (ol@csc.kth.se)

Object recognition with pose detection and object verification in 3-D

For recognizing natural objects in natural scenes in a view-based manner, a natural approach consists of taking a set of images of the given objects from different views. Lowe approached this problem using an affine Hough transform in which local affine approximations to the perspective transformation were made and evidence accumulated for different hypotheses by voting. Object verification was then also made based on an affine approximation of the perspective mapping from each object hypotheses. By the use of existing methods for bundle adjustment, a more refined approach to this problem would be to precompute correspondences between interest points from different views and then build an intermediate object representation of each object that could be used for performing the object verification in 3-D.

The topic of this MSc project is to perform a literature survey of existing methods for object recognition with pose detection and/or object verification in 3-D and then develop a solution to this problem based on the type of relations that can be expressed between image features of each object in the database from a reconstructed sparse 3-D object model refined by bundle adjustment.

Contact: Oskar Linde (ol@csc.kth.se) or Tony Lindeberg (tony@csc.kth.se)