The First Set of Features

  This weekend I'm working on measuring the similarity between training and testing data (actually they are mean value of columns of split odd and even quaters of the image ). As a problem was mentioned that the two sets of data were too similar if I just take the even and odd columns, I have changed it to be even and odd quaters, and get the two sets of data more different as below.

   I have used Matlab to calculate the L2 norm(without being squared)of the training and testing data, and get a histogram as below. It describes the distance matrix.
  First 10 columns describe the distance between the first testing data with all the training data. We can see that the first testing data matches the first training data(the distance is almost zero). And for the second 10 columns, we can get that the second testing data matches the second training data and so on.

  Interestingly, we can see from the histogram that the distance between material 2 and 5 are a little bit too small that we may be confused if we just use this histogram to category them . Lets see the mean value of the columns below, it's really similar. But when we see from the original pictures, we can tell they are totally different(Actually they are paper and metal). So it must be the procedure of getting the average value makes it confusing. As a conclusion, mean values are useful, but not enough.

Some progress in existing data

   After doing some review in parts of the homeworks in cse 252, I'm getting more familiar with Matlab and pattern recognition experiments.So I'm trying to analyse the existing data, split each picture in even and odd columns to form two data set. These pairs of data may help me to determine whether  I'm going in the right direction in selecting the method to solve the problem.(But since the images are too large, I'm getting just 67% of the results)
wood:                                 denim:

foampad:                      foamgrip:

pictures of 4 types of material that are separated

Furthermore, I get some plots of the mean values of the separated images(by column). These values may help me get further features of each material.

wood:                                   denim:             

foamgrip:

mean values of 3 types of material 's columns

Next step, I'll try to distract features from original data and generated plots.

Structured Light in Scattering Media

http://www.cs.cmu.edu/~ILIM/projects/LT/structured/structured.html
We can get some references from this website, which uses Light Stripe to scan object in special media, hence rebuilding the object's 3D model.

Maybe a little bit far away from material recognition, cause it focuses more on 3D rebuilding. But still good to know some related works using light stripe. And 3D rebuilding is also a direction where this project can go. Many things are not decided yet...

http://sites.google.com/site/koppaldev/

About the project

 These are the pictures of the device we have.


Scan the object->get reflection-> analyse it->identify it
->(paper or rubber?)















Problem:
1. We still know little about the scanner.(a little bit awkward)
In this way, we can't decide the project in detail yet, including which kinds of material we are analysing and how we are gonna categorize it.
Solution: Tuesday I'll get more information about the scanner.


2. I'm lack of programming experience in pattern recognization area.
Solution: I'll work hard, and maybe I'm gonna need your help.



Thank you!

Related works

Below is a paper Serge recommended, while it happens to be a paper that demonstrates a possible method that I may be familiar with.
http://physics.fme.vutbr.cz/libs/images/stories/publikace/2001/Lecture%20Notes%20in%20Computer%20Sciences_2001.pdf
(Data below is collected from this paper)


1.WHAT DID THEY GET

LIBS (Laser Induced Breakdown Spectroscopy) spectra is collected for samples of several materials. The picture below shows spectra of (a) aluminium, (b) copper, (c) lead, (d) zinc, (e)mild steel, (f) stainless steel and (g) vitrification glass

Each spectrum collected using a LIBS instrument is a “finger-print“ of the material being analysed and the conditions under which it was collected. 

So we can identify a kind of material by analysing its spectrum.

2.HOW DID THEY ANALYSE IT

Here they used the Mahalanobis Distance method to evaluate the distinction of sample material and the criteria, and then give a discrimination of the material.

The M.Dist applies to all pixels in the spectrum, not just to

one or two pixels of a peak, and therefore can be considered to be a multi-dimensional standard deviation that is applied to the whole spectrum.

I think this may be a possibe solution to our question.

3.WHY IS THIS PROJECT DIFFERENT

In contrast, this is "an example of a project doing material recognition from regular images"  http://people.csail.mit.edu/celiu/CVPR2010/FMD/index.html

It is more relied on a large dataset, but don't need a device. But mine may need to programme on a scanner to get the properties of the material and identify them.

Kinect

Kinect is a motion sensing input device by Microsoft for the Xbox 360 video game console and Windows PCs,  which provide full-body 3D motion capture, facial recognition and voice recognition capabilities.

<><><>

This infrared image shows the laser grid
 Kinect uses to calculate depth.

The depth map is visualized here using color
gradients from white (near) to blue (far)

3D modeling by distance!

                                      From Wikipedia http://en.wikipedia.org/wiki/Kinect

Get started

This weekend, after discussing with the professor, I have decided to do the project  by my own. It's gonna be challenging and interesting. Now I have got the access to learn more about the area i may be interested in.