 |
.. (לתיקייה המכילה) | |
What is the nature of this project, anyway? | |
|
The project is a "real-life engineering task". It means that it isn't completely close, there is no single solution, there are problems, and not all of them are solvable. I'm not trying to test your ability to code and to follow instructions, but to see how you think, your creativity, how well you approach problems and attempt to solve them. Also, not all the methods in the project are meant to do well. I'm trying to show you the real life value of methods, and not the theory. For example, template matching for face detection using SSD is not supposed to work well, but I want you to see it and feel it, analyse it and discuss why. Maybe you can make it work good using some improvements, and maybe not. Creativity beyond the scope of the project will earn you bonus points, but please don't spend countless hours on the project. remember that we have an exam and the project has a deadline. |
In part 1 and 2, can I resize the images using imresize? | |
| Yes, just use the bicubic method. |
How do I build the gradient histograms in part 2? | |
|
Each histogram is composed of 9 bins representing the gradient orientations. The gradient magnitudes will be summed up for each orientation. For example, suppose you have two pixels for orientation 3, and the magnitudes in these pixels are 5 and 7, then the bin for orientation 3 will have the value 12. |
In part 2, should I use only positive values for the gradient directions? | |
|
You may use only positive ([0,pi]) values for the directions. It should bring better results potentially. You can do it by adding pi to the negative angles [-pi,0]. |
Should the recognition with Fisherfaces be better than Eigengfaces? | |
|
Normally it should. However, since we have a small set of examples within each cluster, it may be equal or slightly worse than Eigenfaces. Also, Fisherfaces is much more sensitive to the cropping you did on part 3, so even if these images worked with Eigenfaces they may not work with Fisherfaces. The recognition with faces from the training should work. Try several crops and report your success and conclusions. |
In part5, can we compute the Sb and Sw matrices directly on the samples? | |
|
Yes. Actually, I've been told that it brings better results, even though the theory is the same. The general MIU in the formula will be meanvec, and the other MIUs with notation i are the inner class mean vectors. After computing Sb and Sw this way, you should convert them to the PCA space like this: newS = W ' * S * W, Where W is the pca dictionary computed in the first step. |
Multiplying Wpca with Wfld isn't possible in the way mentioned in the presentation. | |
|
Yes, there is some mix-up there about transpose and stuff. Think about the dimensions of the final dictionary you need, and multiply the matrices accordingly. |
How do I compute the accuracy of the Fisherface method? | |
|
When using c=43, compare to label_test. When using c=21, compare to label_test21. When using c=70, compare to both label_test and label_test70, and add their results. |