SoC 2008 student proposals - Revision history

James Legg: /* James Legg, OpenGL Accelerated Hugin Preview */ added link to project page.

2008-05-06T17:12:15Z

‎James Legg, OpenGL Accelerated Hugin Preview: added link to project page.

Timnugent: /* Tim Nugent, Support Vector Machine-based Sky Identification */

2008-04-09T11:57:23Z

‎Tim Nugent, Support Vector Machine-based Sky Identification

Timnugent: /* Tim Nugent, Support Vector Machine-based Sky Identification */

2008-04-08T17:52:03Z

‎Tim Nugent, Support Vector Machine-based Sky Identification

Timnugent: /* Tim Nugent, Support Vector Machine-based Sky Identification */

2008-04-08T17:51:24Z

‎Tim Nugent, Support Vector Machine-based Sky Identification

Timnugent: /* Tim Nugent, Support Vector Machine-based Sky Identification */

2008-04-08T17:50:37Z

‎Tim Nugent, Support Vector Machine-based Sky Identification

Timnugent: /* Tim Nugent, Support Vector Machine-based Sky Identification */

2008-04-08T17:46:30Z

‎Tim Nugent, Support Vector Machine-based Sky Identification

Timnugent: /* Tim Nugent, Support Vector Machine-based Sky Identification */

2008-04-08T17:45:30Z

‎Tim Nugent, Support Vector Machine-based Sky Identification

Timnugent: /* Tim Nugent, Support Vector Machine-based Sky Identification */

2008-04-08T17:37:50Z

‎Tim Nugent, Support Vector Machine-based Sky Identification

Timnugent: /* Tim Nugent, Support Vector Machine-based Sky Identification */

2008-04-08T17:34:46Z

‎Tim Nugent, Support Vector Machine-based Sky Identification

Timnugent: /* References */

2008-04-08T17:33:10Z

‎References

← Older revision		Revision as of 17:12, 6 May 2008
Line 13:		Line 13:

	== James Legg, OpenGL Accelerated Hugin Preview==		== James Legg, OpenGL Accelerated Hugin Preview==
		+	This project was accepted, see [[SoC 2008 Project OpenGL Preview]] for details.

	* I am James Legg, a 2<sup>nd</sup> year undergraduate in Computer Science and Mathematics at The University of York (UK).		* I am James Legg, a 2<sup>nd</sup> year undergraduate in Computer Science and Mathematics at The University of York (UK).

← Older revision		Revision as of 11:57, 9 April 2008
Line 277:		Line 277:


−	* I am 2nd year PhD Student in the Bioinformatics group at University College London, UK.	+	* I am a 2nd year PhD Student in the Bioinformatics group at University College London, UK.

	* I originally trained as a Pharmacologist, attaining a 2.1 BSc (Hons.) from the University of Bristol. I then completed an MRes (with Merit) in Bioinformatics at Birkbeck College, University of London, before working for a bioinformatics company Inpharmatica Ltd., and then as a bioinformatician for a genetics research project at Queen Mary, University of London. I returned to academia and am currently in my 2nd year of a 4 year PhD course in the bioinformatics group; my project concerns the structure prediction of transmembrane proteins using machine learning approaches (more info here: http://www.cs.ucl.ac.uk/staff/T.Nugent/research.html).		* I originally trained as a Pharmacologist, attaining a 2.1 BSc (Hons.) from the University of Bristol. I then completed an MRes (with Merit) in Bioinformatics at Birkbeck College, University of London, before working for a bioinformatics company Inpharmatica Ltd., and then as a bioinformatician for a genetics research project at Queen Mary, University of London. I returned to academia and am currently in my 2nd year of a 4 year PhD course in the bioinformatics group; my project concerns the structure prediction of transmembrane proteins using machine learning approaches (more info here: http://www.cs.ucl.ac.uk/staff/T.Nugent/research.html).

← Older revision		Revision as of 17:52, 8 April 2008
Line 277:		Line 277:


−	* I am ~~Timothy Nugent, a~~ 2nd year PhD Student in the Bioinformatics group at University College London, UK.	+	* I am 2nd year PhD Student in the Bioinformatics group at University College London, UK.

	* I originally trained as a Pharmacologist, attaining a 2.1 BSc (Hons.) from the University of Bristol. I then completed an MRes (with Merit) in Bioinformatics at Birkbeck College, University of London, before working for a bioinformatics company Inpharmatica Ltd., and then as a bioinformatician for a genetics research project at Queen Mary, University of London. I returned to academia and am currently in my 2nd year of a 4 year PhD course in the bioinformatics group; my project concerns the structure prediction of transmembrane proteins using machine learning approaches (more info here: http://www.cs.ucl.ac.uk/staff/T.Nugent/research.html).		* I originally trained as a Pharmacologist, attaining a 2.1 BSc (Hons.) from the University of Bristol. I then completed an MRes (with Merit) in Bioinformatics at Birkbeck College, University of London, before working for a bioinformatics company Inpharmatica Ltd., and then as a bioinformatician for a genetics research project at Queen Mary, University of London. I returned to academia and am currently in my 2nd year of a 4 year PhD course in the bioinformatics group; my project concerns the structure prediction of transmembrane proteins using machine learning approaches (more info here: http://www.cs.ucl.ac.uk/staff/T.Nugent/research.html).

@@ Line 312: / Line 312: @@
 ) Write standalone tool in C++ to score all regions of the image, and return a posterior probability of whether or not that region contains cloud. Create a scoring function which analyses the results globaly - e.g. Score up 'cloud' regions that are in contact with a large number of other 'cloud' regions, and score down those that are not. Use an adjustable threshold to trim control points from associated Panotools/Hugin file. This should take about 3 weeks.
 ) Write a GUI based version using wxWidgets/GTK etc. This should take about two weeks. I see this  as the least important component of the project as I envisage the tool being run from the command line, or via Hugin, so should other stages of the project fall behind schedule, this part would be sacrificed.
 ) Testing, both code (across multiple platforms) and prediction accuracy. This will take 2 weeks but will begin during (5).

@@ Line 297: / Line 297: @@
 * My preliminary plan for the project is as follows, beginning immediately:
-. Familiarise myself with:
-*Hugin source code - begun, will continue for duration of the project.
+) Familiarise myself with:
-*VIGRA image processing library - begun, will continue for duration of the project.
+# Hugin source code - begun, will continue for duration of the project.
-*libSVM/SVMlight source - already done to a large degree.
+# VIGRA image processing library - begun, will continue for duration of the project.
-*FlickR API - begun.
+# libSVM/SVMlight source - already done to a large degree.
-*Feature/texture classification algorithms (e.g. Gabor filtering).
+# FlickR API - begun.
-. Assemble training and test sets of images containing/not containing cloud, probably from community image sites such as Flickr, using FlickR API/Perl scripts. This will be ongoing in order to obtain maximum diversity, but in the first instance it should take approximately two weeks to collect and then accurately classify enough images (and the regions within them) to begin SVM training.
+) Assemble training and test sets of images containing/not containing cloud, probably from community image sites such as Flickr, using FlickR API/Perl scripts. This will be ongoing in order to obtain maximum diversity, but in the first instance it should take approximately two weeks to collect and then accurately classify enough images (and the regions within them) to begin SVM training.
-−
-. Use VIGRA, and perhaps other approaches, to extract image features from all regions (e.g. RGB data, smoothness, edges etc - motifs common to cloud textures). Add additional global features, for example the (x,y) co-ordinates of the region within the image). Finding the features that accurately discriminate between the two sets may not be straightforward so I envisage this will be the hardest part of the project and may take 2 months.
-−
-. Use these features to train an SVM classifier. Optimise SVM parameters (possibly using genetic algorithm based grid search), cross-validate, re-train, optimise (and repeat..). Use UCL Linux cluster to run many SVM training jobs in parallel. This stage will overlap with (3) to a large degree, finishing  two weeks afterwards.
-. Write standalone tool in C++ to score all regions of the image, and return a posterior probability of whether or not that region contains cloud. Create a scoring function which analyses the results globaly - e.g. Score up 'cloud' regions that are in contact with a large number of other 'cloud' regions, and score down those that are not. Use an adjustable threshold to trim control points from associated Panotools/Hugin file. This should take about 3 weeks.
+) Use VIGRA, and perhaps other approaches, to extract image features from all regions (e.g. RGB data, smoothness, edges etc - motifs common to cloud textures). Add additional global features, for example the (x,y) co-ordinates of the region within the image). Finding the features that accurately discriminate between the two sets may not be straightforward so I envisage this will be the hardest part of the project and may take 2 months.
-. Write a GUI based version using wxWidgets/GTK etc. This should take about two weeks. I see this  as the least important component of the project as I envisage the tool being run from the command line, or via Hugin, so should other stages of the project fall behind schedule, this part would be sacrificed.
+) Use these features to train an SVM classifier. Optimise SVM parameters (possibly using genetic algorithm based grid search), cross-validate, re-train, optimise (and repeat..). Use UCL Linux cluster to run many SVM training jobs in parallel. This stage will overlap with (3) to a large degree, finishing  two weeks afterwards.
-. Testing, both code (across multiple platforms) and prediction accuracy. This will take 2 weeks but will begin during (5).
+) Write standalone tool in C++ to score all regions of the image, and return a posterior probability of whether or not that region contains cloud. Create a scoring function which analyses the results globaly - e.g. Score up 'cloud' regions that are in contact with a large number of other 'cloud' regions, and score down those that are not. Use an adjustable threshold to trim control points from associated Panotools/Hugin file. This should take about 3 weeks.
 [[Category:Community:Project]]

← Older revision		Revision as of 17:46, 8 April 2008
Line 272:		Line 272:

	[[Category:Community:Project]]		[[Category:Community:Project]]
		+

	== Tim Nugent, Support Vector Machine-based Sky Identification ==		== Tim Nugent, Support Vector Machine-based Sky Identification ==

@@ Line 280: / Line 280: @@
 * I originally trained as a Pharmacologist, attaining a 2.1 BSc (Hons.) from the University of Bristol. I then completed an MRes (with Merit) in Bioinformatics at Birkbeck College, University of London, before working for a bioinformatics company Inpharmatica Ltd., and then as a bioinformatician for a genetics research project at Queen Mary, University of London. I returned to academia and am currently in my 2nd year of a 4 year PhD course in the bioinformatics group; my project concerns the structure prediction of transmembrane proteins using machine learning approaches (more info here: http://www.cs.ucl.ac.uk/staff/T.Nugent/research.html).
-* I have programmed in C/C++ for over two years and have recently implemented a dynamic programming algorithm in C++ which is used in protein structure prediction. I have submitted patches for Hugin, and am currently working on a game that used the SDL libraries.
+* I have programmed in C/C++ for over two years and have recently implemented support vector machine (SVM) and dynamic programming algorithms in C++ into a protein structure prediction program. I have submitted various patches for Hugin, and am currently working on a game that used the SDL libraries.
-* I have a strong interest in photography, particularly panoramic and underwater. I am particularly keen to combine these two areas as there are very few examples where they overlap. I'm familiar with a range of stitching tools, but tend to use Hugin as I rarely work on anything other than a Linux box. Here are a few examples:
+* I have a strong interest in photography, particularly panoramic and underwater. I am particularly keen to combine these two areas as there are very few examples where they overlap. I'm familiar with a range of stitching tools, but tend to use Hugin as I rarely work on anything other than a Linux box. Here are a few examples: http://www.cs.ucl.ac.uk/staff/T.Nugent/panoramics/
 http://www.cs.ucl.ac.uk/staff/T.Nugent/panoramics/
 * So far I have not used Hugin/Panotools for anything other than stitching panoramas. However, I'm keen to experiment with HDR photography.
-* I'm a developer for the Bioperl project (http://www.bioperl.org) which produces opensource Perl code for the bioinformatics community. You can download the development branch via SVN, which includes some of my modules for drawing graphics using libgd and SVG are available here. I intend to port these to Python for the increasingly popular BioPython project.
+* I'm a developer for the Bioperl project (http://www.bioperl.org) which produces opensource Perl code for the bioinformatics community. You can download the development branch via SVN, which includes some of my modules for drawing graphics using libgd and SVG are available here. I intend to port these to Python for the increasingly popular BioPython project. http://www.cs.ucl.ac.uk/staff/T.Nugent/code.html
 http://www.cs.ucl.ac.uk/staff/T.Nugent/code.html
-* The reason I've chosen to tackle the sky identification problem is that I'd like to apply my knowledge of support vector machine (SVM) classification to a novel problem outside biology. Sky identification ties in perfectly with my interest in panoramic photography, and as someone who has had to manually remove control points from rogue clouds, I feel I'm the ideal person to attack this problem. SVM classifiers have been used in a range of areas from finance to face recognition and there is huge scope for applying these tools to new fields. I've spent the first half of my PhD familiarising myself with the SVM algorithm, various software packages, parameter optimisation and methods of cross-validation for applying SVMs to protein structure prediction. Initial results indicate SVMs are more accurate than the other two most popular methods used in this area - neural networks and hidden Markov models; however different tools will undoubtedly perform better in some areas but worse in others, so it must not be taken for granted that SVMs will be the optimal solution for the sky identification problem. Training and cross validating SVMs can be computationally expensive. I'm fortunate to have access to a large Linux cluster at UCL running Sun Grid Engine, which I intend to use for this project (http://www.ucl.ac.uk/media/library/Dell).
+* The reason I've chosen to tackle the sky identification problem is that I'd like to apply my knowledge of support vector machine (SVM) classification to a novel problem outside biology. Sky identification ties in perfectly with my interest in panoramic photography, and as someone who has had to manually remove control points from rogue clouds, I feel I'm the ideal person to attack this problem. SVM classifiers have been used in a range of areas from finance to face recognition and there is huge scope for applying these tools to new fields. However, training and cross validating SVMs can be computationally expensive. I'm fortunate to have access to a large Linux cluster at UCL running Sun Grid Engine, which I intend to use for this project (http://www.ucl.ac.uk/media/library/Dell).
-For this implementation I intend  to train an SVM classifier to automatically and accurately identify regions in an image that contain cloud, and those that do not, so that control points can be removed or excluded therefore increasing the quality of the alignment. This may also reduce the computation time for control point creation, by reducing the available regions within which to search for control points. This process could also be extended to other regions within an image that do not remain static over a series of photos, e.g. water, waves, other windswept regions etc. I will develop a standalone tool in C++ to score all regions of the image, and return a posterior probability of whether or not that region contains cloud. A scoring function and an adjustable threshold will be used to trim control points from associated Panotools/Hugin file, or create a mask with which to exclude control point searches. I also intend to produce a  GUI based version using wxWidgets/GTK etc. The source code will be freely available, and I will provide binaries for all the major platforms.
+* For this implementation I intend  to train an SVM classifier to automatically and accurately identify regions in an image that contain cloud, and those that do not, so that control points can be removed or excluded therefore increasing the quality of the alignment. This may also reduce the computation time for control point creation, by reducing the available regions within which to search for control points. This process could also be extended to other regions within an image that do not remain static over a series of photos, e.g. water, waves, other windswept regions etc. I will develop a standalone tool in C++ to score all regions of the image, and return a posterior probability of whether or not that region contains cloud. A scoring function and an adjustable threshold will be used to trim control points from associated Panotools/Hugin file, or create a mask with which to exclude control point searches. I also intend to produce a  GUI based version using wxWidgets/GTK etc. The source code will be freely available, and I will provide binaries for all the major platforms.
 * As a full time PhD student, I do not have a summer break as undergraduates do, therefore am unable to contribute a 40 hour week to the project. I intend to use any free time I have inside or outside university hours, and envisage working on the project for most evening during the week and for a significant portion of the weekend. While this may not be ideal, I feel the expertise I already have will allow me to 'hit the ground running'. I am also able to start the project immediately so I am highly confident I can finish it by Google's 'pencils down' date. Should I encounter problems, I'm willing to take my annual leave in order to work full time on the project.
 * My preliminary plan for the project is as follows, beginning immediately:
 . Familiarise myself with:
-**Hugin source code - begun, will continue for duration of the project.
+*Hugin source code - begun, will continue for duration of the project.
-**VIGRA image processing library - begun, will continue for duration of the project.
+*VIGRA image processing library - begun, will continue for duration of the project.
-**libSVM/SVMlight source - already done to a large degree.
+*libSVM/SVMlight source - already done to a large degree.
-**FlickR API - begun.
+*FlickR API - begun.
-**Feature/texture classification algorithms.
+*Feature/texture classification algorithms (e.g. Gabor filtering).
 . Assemble training and test sets of images containing/not containing cloud, probably from community image sites such as Flickr, using FlickR API/Perl scripts. This will be ongoing in order to obtain maximum diversity, but in the first instance it should take approximately two weeks to collect and then accurately classify enough images (and the regions within them) to begin SVM training.
@@ Line 317: / Line 314: @@
 . Testing, both code (across multiple platforms) and prediction accuracy. This will take 2 weeks but will begin during (5).
 [[Category:Community:Project]]

@@ Line 270: / Line 270: @@
 # OpenRaster, http://create.freedesktop.org/specs/OpenRaster-draft.pdf
 # Personal website, http://cs.bilkent.edu.tr/~onurk/
 [[Category:Community:Project]]