Project 4 Artifact
by Oscar Danielsson
General Remarks
- The interface for the std::list that we used for this project was not very intuitive. For example, all descriptions that I read said that when inserting an element into the list using the insert(iterator,element) function, the new element will be inserted after the element to which the iterator points. This turned out to be wrong and in fact the new element is inserted before the element to which the iterator points. This took a lot of time to discover.
- Another list related problem was that the provided sort algorithm did not work and the ordering out put from the algorithm did not put the eigenvalues in descending order (not in ascending order either for that matter). That resulted in worse recognition performance than expected. (maybe the list was behaving strangely because I was working in VC++ 2003?)
- The list implementations obviously differ between VC++2003 and VC++2005 – for example: I used dynamically allocated Position structures in my list and in VC++2003 it worked well to use the delete command to free that memory space when popping or removing elements from the list. In VC++2005, however, the delete command caused the program to crash.
- In order to make sure the algorithm finds the faces in the images, one has to make sure to go through big enough scales (bigger scales mean that the 25 by 25 face finding window encapsulates less of the searched image). The reason is that in order to get a good mach the face finding window needs to include only the center of the face (that is the mouth, nose and eyes), but nothing else.
- Finally, it turned out that dividing by the standard deviation instead of the variance when calculating the error in the findFace routine gave slightly better performance.
Extra Credit
- I implemented the speedup for 1 whistle.
Eigenfaces and Average Face
Below are the ten most significant eigenfaces along with the average face.
Average Face
Eigenfaces
Recognition Results
The figure below shows the results when using a varying number of eigenfaces acquired from the set of images of nonsmiling students to recognize the images of the smiling students.
It is obvious that one gains little by using more than 10-11 eigenfaces. In some cases the algorithm did not do correct classification for any number of eigenfaces. A couple of examples of hard cases are shown below:
Image 01 was often
recognized as image 22:
à
Image 13 was often
recognized as image 04:
à
Finding Faces in Images
Using ten eigenfaces to span face space we shall try to find patches representing faces in images. Several examples are shown below.
Cropping elf.tga
(scales: 0.35 – 0.45, step: 0.01)
Original
Cropped
Cropping a picture of
myself (scales: 0.15 – 0.25, step:0.01)
Original (scaled down)
Cropped
Finding the three
faces in IMG_0031.tga (scales: 0.45 – 0.55, step: 0.01)
Original
Faces marked
Finding faces in
another group photo (scales: 0.7 – 1.2, step: 0.05)
Original
Faces marked
We see that the algorithm can have problems with images where the background has similar colors and shapes as the faces. We can also see that the algorithm does not find faces that are partially occluded, which was expected. Getting good performance out of the algorithm and finding the faces in an image usually comes down to searching a good range of scales. That range differs a lot between different images and often some intervention from the user is required to find good scales. If we had unlimited time or processing power, we could off course get around this problem by just searching a wide range of scales with a small step size.
Batch files: