1. Stereo reconstruction is only useful for reconstructing a simple scene from two nearby
views. It can be difficult to extend to complicated scenes with occlusion and to a large
number of cameras or views. Volumetric reconstruction is one way to handle a large
number of cameras with arbitrary viewpoints. In this assignment you will be required
to fuse the data from 36 viewpoints of an object into a detailed volumetric model of
that object. An example result is given below.
36 viewpoints of a toy dinosaur spinning on a turntable (10° per image).
36 projection matrices which determine how each spatial point maps onto the image
A volumetric model of the dinosaur
(To be submitted) 4 new viewpoints of the reconstructed dinosaur which do not
coincide with any of the original 36 viewpoints. Try to texture map the images.
The method you are required to implement is known as “shape from silhouette”. This has
been covered in lectures. Images and some hints are available from Blackboard as dino.zip.
2. Build a face recognition system based on the eigenface technique to recognize faces
from the database on the website. You should be able to obtain about 96% correct
recognition. Supply the code listing, explain your method, and performance analysis.
Higher marks can be obtained by building a nice gui interface in Matlab. Face images are
available from Blackboard. Details of the Eigenface method are in the book chapter “Face
Recognition for Data Mining” on Blackboard.
(Total 20 Marks)