face detection and recognition

Alan Chatt, ajc201@ecs
David Johnston, dj301@ecs
Kevin Lau, kl1101@ecs
Ben Tindall, bt101@ecs

Download Training Set (400kb)

How to use the eigenface applet

Requirements

Instructions

  1. Install the Java JRE and ensure this is used by the browser
  2. Install J3D libraries
  3. Using the policy tool supplied with the Java JRE grant permission for the applet at the address below to read a location on your hard drive where the training data will be unzipped to.
  4. Download and unzip the demonstration data.
  5. Use the load Images button to locate the folder containing the training data ('data'). This method assumes each sub folder contains the training images for that person.
  6. Click the crop faces button to have the applet detect the faces and crop the images to them.
  7. Now click train to compute the eigenfaces and features vectors for the training set.
  8. The top three dimensions in the feature space can be visualised by clicking the 'Display Feature Space' button. This opens a separate window with a 3D graph which can be manipulated by click ‘n dragging the mouse. (left to rotate, middle to zoom, right to move).
  9. To identify a face there is a sub folder named 'lookup' within the demo data. This contains images not used in the training set. Click the identify face and locate a face in this subset. When the face is loaded, it will be cropped automatically. The applet will then compute the Euclidian distance from each face in the training set, and use the K nearest neighbour method to determine the classification of this new face. The training set is re-ordered by distance from the probe face, and the background colour faded in proportion to the distance from the face. The position of this face can be visualised in the feature space by clicking the 'Display Feature Space' space button again. It will appear as a single cross.

Important: Always remember to `crop faces' before running identification. Otherwise the system will compare the uncropped faces against the (cropped) face that is loaded.

Credits

more information