Career Developments

My research unit (RSRU) is currently setting up our own MODIS base station. Very interesting indeed.

We are planning to implement a number of algorithms for research purposes and would therefore need a few developers to do the dirty work. I might be coming on-board to help as soon as I’ve finished my thesis.

Let’s hope everything works out!

Thesis Progress

I’ve been spending a lot of time on the literature review, which has taken much longer than I thought. As I progressed, I found new articles which I felt had to be summarized. In the end I had around 90-100 articles to review. I’ve reviewed around 30-40 articles already. The workload became too much / dull, so I decided to give it a bit of a break and move onto the other chapters.

I’ve been restructuring the document a bit and adding information here and there.The document has a bit more volume and is starting to look like a thesis. I have 92 pages at present. There’s still a lot of work left though. I just need to keep my focus in these final few weeks…

Thesis Progress – Literature Review

I’ve been trying really hard to keep my focus, but it is hard reading research articles from 9-5.

I’m almost half-way with my literature review. The language is very sloppy at the moment and I would need to go back to clean that up and to work on the flow a bit. My goal is just to get through all the papers and get them summarised as soon as possible.

I still have around 50 papers to go, so it’s not going to be easy. Luckily I’m just updating an existing literature study I did in 2006, but 50 articles is a lot!

Wish me luck. I’m going to need it.

Thesis Layout

I had some corrections to make to my paper. I’ve just finished the second draft and waiting to hear back from my supervisors.

In the mean time, I’ll continue finalizing my thesis layout. Luckily I’ve already done a fair bit of writing, so the literature review is well on its way and I have two papers I can use in my experiments section.

Let’s hope I finish my first draft before the end of Feb. Its sure going to be tough!!

Gap Breaching

We had a look at our results and noticed that there were some gaps in a number of our ouptuts. It was decided to try and breach these gaps by converting our raster centerlines into vector format.With the help of the open source GIS application GRASS we were able to do the conversion and also do some elementary clean-up work on the vectors. This included snapping of nearby vectors, smoothing/straightening of wiggly lines and removal of small anomalies. It turns out that GRASS has a module (developed as part of Google’s Summer of Code 2007) that will do various clean-up operations that are specific to road networks. There’s even a tutorial that explains the how some of the components works. Very handy!

Based on our new clean set of vectors, we implemented a very basic algorithm that breaches gaps based on the following criteria:

• the orientation of endpoints,
• the distance between endpoints, and
• the length of the line segments to be connected.

Our method works really well and it performs as expected. It connects a lot of the lines I would connect, if I did not know what the input data looked like.

Unfortunately some of our results include a lot of false positives. The gap breaching algorithm connects these false positives (FP) and we end up with even more incorrect line segments.

This is just a rough guess, but I would say that we see improvements of around 5-15% on data sets with little FPs and a couple of gaps. Unfortunately the real world data sets contains FPs and we are unable to improve on our highest quality measures of 64%. Bummer.

The next step would be to reduce the number of FPs, which obviously causes the number of true positives (TP) to drop. Hopefully the gap breaching algorithm will be able to compensate for the loss of TPs, but I’m not holding my breath.

Paper Results

I’ve completed my paper and is in the process of submission. Hopefully the paper will pass through the peer review unscathed, but I suppose few are that lucky.

Since it will be a while before the paper is in print, I shall give a sneak informal preview of the results.

In the paper, we investigated two matters:

1. Image processing is typically done on a pixel-by-pixel basis. Recently, a number of approaches have opted to segment an image and process each segment as a whole. We investigated whether there is any gain in accuracy when adding a segmentation component into our road extraction system.
2. Some road extraction systems employs several classifier to identify road pixels/segments. The outputs from these classifiers is then fused in order to provide more consistent and accurate results. Again we investigate whether this holds for our system.

The input image and ground truth:

Input Image	Ground Truth

1. Segmentation vs. pixel based approaches:

Here is the results from the segment based approach:

Classified Image	Thresholded Result	SORM Output

Here is the results from the pixel based approach:

Classified Image	Thresholded Result	SORM Output

The outcome was that in this case (using the fusion of classifiers) the segmented approach (S) provided better results than the pixel based approach (P). S achieved a quality measurement of 60.7% whilst P managed 51.5%. Furthermore, the standard deviation for S was lower at 15.4% than P’s 18.8%.

2. Single classifier vs. data fusion for pixel and segmented approaches:
For the second experiment, we removed the fusion component and tested each classifier separately using the pixel and segmented approaches.

For P, the classifier using the Mahalanobis distance function achieved the highest quality. Here is some sample images:

Classified Image	Thresholded Result	SORM Output *

For S, the classifier using the Bhattacharrya distance function achieved the highest quality. Here is some sample images:

Classified Image	Thresholded Result	SORM Output *

* Please note that the images above for the Mahalanobis pixel based approach (PM) and Bhattacharyya segment based approach (SB) might not be representative of the entire data set we used.

The results showed that PM achieved an quality of 64.6% with a 17.3% standard deviation, opposed to SB’s 62.8% and 13.6%.

One can therefore not conclude that S will always have a higher quality measurement, but does definitely prove to be more consistent than P.

Furthermore, data fusion reduced the quality of our results and we suspect that it is because the problem only has two classes (road and non-road). The classifiers are constantly in direct conflict/agreement of the result. We suspect data fusion to be more suited for applications where say sensor 1 detected an object as being A or B and sensor 2 detected and the same object as being B or C. The fusion of this result should yield the correct answer of B.

Our next step would be to create a higher level system to breach the gaps that exist in the results. We hope to raise our quality measurements above 70%.

Thesis Writing Tips

My research supervisor gave us all these general and practical research tips. I thought it should be quite relevant to everyone busy with postgrad studies.

1. Do not make use of informal language. This is an academic work and needs to be written formally. What I mean by this is do not use “we”, “I”, “they”. Write passively.

2. Experimental work and conclusions are written in past tense. The rest in present tense.

3. Refer to everything by name. Do not make use of “it”, “this”, “these”, unless there is absolutely no ambiguities (how ever you spell it).

4. Try not to write long sentence. Writing should be like coding. One concept per sentence.

5. Be sure to write with flow. Each sentence/paragraph/section/chapter should follow on the previous. The reader should at all times know what he/she is reading and why.

6. If you have many mathematical symbols and acronyms, define them in appendices. This will also help you to make sure that you assign one meaning to a symbol. Do not overload the meaning of symbols. It makes reading difficult. Do not use different symbols for the same meaning.

7. Try not to use quotes and footnotes. Footnotes break the flow. Quotes say what someone else think. The purpose of the thesis is to show what you think.

8. Your background should be as complete as possible. You need to have in the background all theory, previous models, etc that have relevance to what you are doing. And remember, do not assume that the reader has prior knowledge. You need to write to show that you understand the field, that you can critically discuss and evaluate existing literature. Do not assume that the reader will read between the lines.

9. The introduction chapter is very important. It provides the vehicle through which you set the stage. It is the overture of an opera. You need to position your work, and motivate why you have done this work. The introduction states your objectives and how these objectives will be addressed. You outline your contributions. Then you tell the reader what to expect for the rest of the thesis by giving an outline of the rest of the thesis. The reader should, after reading the introduction have a precise idea of what you are doing, why, how and where.

10. The conclusions chapter is just as important. Herein you state what the objectives were, how they have been addressed and what were the main findings. Then you give ideas of future research emanating from yours.

11. Each chapter should end in a conclusions section, where you summarize the objectives of the chapter, how these have been achieved and what the main findings were. Then you introduce the next chapter and relates it to the current one — to conserve flow.

12. Each chapter should start with a short paragraph just to link it with the previous chapters and to state the goals of the chapter. This comes before the Introduction section of the chapter where you elaborate more.

13. I get many questions as to the length of a thesis. This is difficult to say, because it depends on the type of thesis. My believe is that you write what is necessary, and that’s that. However, good guidelines are that a Masters thesis is approximately 130-200 pages, and a PhD 200-300 pages. But again, when you have written all than you can say in less (or more), then that is your thesis.

14. Make sure that you have up-to-date references. You need to make sure that you obtain and read the recent literature. Also, all references must be cited in the thesis. References must be complete, and in consistent format. Cite original references. For example, if you make use of backpropagation, cite the original work of Werbos.

15. Before you start writing, plan the thesis. Set up a table of contents and plan what you expect to include in each chapter and section. If you have a plan like this, send it to me for comments. It will also be a good idea to look at other theses. Download those from the CIRG website.

16. All figures and tables need to be discussed and analyzed.

17. Use a spell-checker!

Extended Classifier

I’ve finished extending the Texture Cube classifier to work on a segmented image. The preliminary results looks good. However, based on visual inspection, one can never quite tell whether a certain method is better or not. What is pleasing to the eye, might not be the best thing for the machine.

Input image	Single pixel texture cube classifier	Segment based texture cube classifier

I’m busy adding the new segment based texture cube classifier into my image processing chain. Once that is finished I can re-run all my experiments and add the results to my paper.

As mentioned previously, I shall post some of the results on this site.

Paper Nearing Completion

It has been a while since my previous post. The paper is nearing completion. There has been some hitches along the way and took much longer than expected, but has been a brilliant learning experience.

I realized, along the way, that there were some gaps in my approach; not that the approach was incorrect, I just had to investigate some other avenues to be able to defend my results.

The bottom line: I need to extend one of my classifiers, run the experiments and insert the results into my paper.

After that’s finished, I’ll ad some more results. We have been able to raise extraction accuracy to 64%, which is not a massive improvement, but every little bit helps.

Some Results

Here are some of the results as promised.

The basic process is to:
1) classify the image with several classifiers,
2) fuse the results from the different classifiers, and
3) extract centerlines with SORM [1].

It worked quite well for this scene (click to enlarge):

Input Image	Classified Image	Centerlines

And really struggled on this image:

Input Image	Classified Image	Centerlines

This would be an example of the average result:

Input Image	Classified Image	Centerline

At the end of the day, the method managed to achieve around 63.3% accuracy with a standard deviation of 15.3%.

These results will be compiled into a journal article. I’ll provide more specific details as time progress.

References:
[1] P. Doucette, P. Agouris, and A. Stefanidis, “Automated road extraction from high resolution multispectral imagery,” Photogrammetric Engineering & Remote Sensing, vol. 70, pp. 1405–1416, Dec. 2004.