Automatic Landmark Generation

Introduction

Many object recognition methods use flexible models of the objects' shape in order to locate new examples of the object. Object shape models are often constructed by looking at a number of training examples of the object. The shape of each training example can be described by a number of landmarks, each of which represents exactly the same point in all training examples. The top left and top right images show two training examples of faces which have been labelled by hand with common landmarks.

The landmarks are normally hand placed, which is a very laborious and time-consuming task. Badly positioned landmarks can effect how successfully the model can be used to locate new examples of the object. This project is looking at ways in which an optimum set of landmarks can be generated automatically.

Locating Salient Points in a Single Image

Initially we looked at locating points in a single image of the object which are significantly different to all other points in that image. These points are known as salient points. Because salient points are not easily confused with other points within the object, they have a higher probability of successfully being relocated in new images containing examples of the object. Thus salient points should make good landmark points.

The bottom left image shows a saliency image. This indicates the saliency of each area of the face. White areas are the most salient and dark areas are the least. The bottom right image is the original image of the face with the most salient points marked. These coincide with landmark points which might be positioned intuitively .

Locating Salient Points in Multiple Images

Locating salient points in a single image has a number of problems. Firstly salient features which happen to be obscured in the single image will not be considered (eg, eyes being obscured by glasses). Secondly, points which are not a part of the object could be picked as being salient (eg, points on the glasses) and finally the success to
which a point can be relocated is not only dependent on how unique it is but also how much it varies between examples of the object. To solve these problems we have looked at locating salient points by considering a number of training examples.

We have shown that salient points can be relocated in similar images with a higher degree of success than hand chosen landmarks.

Future Work

The objective of this project is to provide software which will automatically generate landmarks on a set of images with minimum user intervention. It is anticipated that this will improve the usefulness of any models trained using these landmarks.

The following lists a number of areas which the project will cover in order to meet its objective.

Modelling the relative positions of a set of landmarks.
When searching for a set of landmarks in a new image, we can use a model to force the landmark matches to form sensible shapes, ie landmarks on the nose should lie somewhere between the two eyes and the mouth.

Comparing Feature Detectors. Feature Detectors allow an area of an image to be described as a vector. There are many different type of feature detectors. Because we use feature detectors to locate salient points it is important to understand the advantages of each kind.

New applications. We intend to apply this work to applications other than labelling faces such as labelling various medical images.

Acknowledgements

Funding for this PhD project was provided by the EPSRC.

Kevin Walker:

Research Courses Clinical Radiology Industrial liaison Contact Us Join Us Search Home

Top Page