Artiste News Archive

Aug 2002

Project completed and evaluated by the EU.

June 2002

Week Beginning 17.06.02

• 17.6.02 Webpages: Got together the various fvg components into tar balls and stuffed them up here. I'll try to keep them up-to-date. Note, that the MPWT may not be working due to ongoing work with the Multiscale feature. The other multiscale features will be unaffected. I'm having trouble generating a statically linked version of fvg at the moment.
• 17.6.02 Web pages: Moved the previous news and put it on the news page

May 2002

Week Beginning 13.04.02

• 13.5.02 General: Well, today's the only day I'm back this week, because tomorrow we fly out to Florence for the PMT meeting in Giunti Labs. Been getting the talk ready for that.
• 13.5.02 Web pages: Moved and updated the previous news and put it on the news page

April 2002

Week Beginning 29.04.02

• 29.4.02 General: From tomorrow I will be away until 13th May 2002.
• 29.4.02 General: Packaged up and sent to IT Innovations the latest MPWT code - including the new generic Multiscale parser, the PWT, and MPWT feature vectors and image processors, and some incidental code to go along with it.
• 29.4.02 General: I've packaged up and sent to IT Innovations the latest module. Query by Fax was sent to IT Innovations at midday today. Also I have sent on the updates to package 4 (the colour matcher code) at the same time.
• 29.4.02 Web pages: Moved last week's news to the news page

Week Beginning 22.04.02

• 26.4.02 General: At this point, I was going to be off for the next 2 weeks, however, I'll now be around next week on Monday only.
• 26.4.02 General: I've been busy with Faizel this afternoon attempting to get the PWTretrieval code working. We've also had a look at the MPWT code, which seems to be working fine. We are still unsure why the PWTretrieval is giving dodgy results. However, this all means I've not had chance to package up the QbF code to get to IT Innovations. I'm pretty sure it's working ok, but it would be nice to test it again.
• 26.4.02 Code: Well, a bug has been found in PWTretrieval which caused the whole thing to bomb out. I found it was due to an array overrun, that's always been there!! Faizal has a newer version of PWT with this sorted, however, it doesn't seem to give correct results. So he's looking into this now.
• 26.4.02 Code: I found a bug in ImageProcessor::GetRGBImage that a VImage was being passed back from a function, which would have caused it to go out of scope. For this reason I've changed it to a VImage *, allocated with new, which must be deleted afterwards.
• 26.4.02 Code: I've got the query by fax module working, and it gives the same results as Faizal got with our test set, so that can be considered working. I will now check it for memory leaks, and send it off to IT Innovations later today. The Multiscale PWT's results are, as I've said, less convincing and this may be left until the beginning of next week, which means one day out of my PhD time.
• 25.4.02 Code: I've tested the query by fax module in the context of fvg. There was a problem, but it was to do with the ReadHeaderFromMemory function returning the wrong value. That's fixed and the module gives results. I'm now generating the query by fax features for the FaxDemo set (which is the same as the VAMDemo set), so that we can actually try to see if it's better.
• 25.4.02 Code: Miles and I are now breaking every rule in the book and developing in the same directory. Miles is installing the normalisation code into the modules for which we have normalisation values. I'm checking the PWT. The results aren't all that convincing, to be honest, but I do believe it's working correctly. The QBF is the next one to test get working and test!
• 25.4.02 Code: I've been sorting out the MPWT code today, in both fvg and Multiscale programs. I've found some bugs in both Multiscale and the MPWT code, so I've been sorting them out. Unfortunately one was really hard to track down. All it was, is that the Multiscale (feature) code wasn't fclose()ing a feature vector which it opened!
• 24.4.02 Code: To work with the MPWT feature (in fact any multiscale feature) requires that the query feature be a single scale version of the multiscale feature. To achieve this I had to play around with fvg so that now there's a very slightly altered function convertNameToMS, that now takes a type for query feature and reference feature. Care also needs to be taken when writing a multiscale feature using the generic multiscale code to pass the multiscale code the correct type of object (i.e. single scale objects), otherwise you get a nasty indefinite loop thing going.
• 24.4.02 Code: A quick proof to show that the generic multiscale code is nice and easy to work with, is that I've just written a multiscale version of a single scale algorithm in under an hour. Got to go see if it works now!
• 24.4.02 Code: I decided it was high time to actually start work on the multiscale PWT because up til now I've been working on the generic Multiscale module, which, although close, does not conform to the API. So I have to write separate modules which are API compliant and use the generic multiscale code for their multiscale implementation.
• 24.4.02 Code: I've updated fvg so that now it is able to accept extra options into module code (such as histogram sizes or whatever) quite easily. This just requires that a special function is written to deal with your module's special options.
• 24.4.02 Webpages: Updated Steps to implementing an Artiste API compliant feature matcher to show how to integrate a feature vector with the fvg code.
• 23.4.02 Code: At the end of the day, I've managed to complete the new version of fvx which to differentiate I've called fvg (feature vector generator). It takes similar commandline as fvx, but now allows comparison of multiple features. The main improvement, though, is the size difference and the ease of use. fvx was 1378 lines of code, at last compile. Although fvg hasn't yet got all of the feature vectors in (which take about 5 lines of code each), fvg is only 296 lines of code. The fact that each new vector takes such a small amount of new code, makes the whole thing easier to use. I needed to do this upgrade to test the new modules we're getting ready for IT Innovations at the end of this week.
• 23.4.02 Code: Today, I've also re-written the ubiquitous fvx which is now getting rather long in the tooth. The main problem with it was its inability to work with any type of FeatureVector. Although it's not finished the code is going to be able to do this. This neans the generate/compare code is generic.
• 23.4.02 Code: Been sorting out the Multiscale code today. I found a bug in the PWT code which Faizal gave me which was causing the Multiscale code to trip over (it was overwriting a loop counter). The Multiscale code is still a bit ropey itself, but it loads and saves vectors ok now, although I'm not sure if the comparison code is working correctly at the moment. It doesn't yet do the compression as standard as that code is not finished.
• 22.4.02 Code: Temporarily, I've gone back to using statically linked libraries so that I can work on the Multiscale PWT and the query by fax. I will re-investigate the problems I was having with shared libraries after the April deadline.
• 22.4.02 Code: The Query by fax algorithm has been integrated into the abstract base class API version. It doesn't quite conform to the API, because of an oversight in the API design where it is assumed the query feature and the reference feature would be built the same way. However, the query by fax algorithm requires a different feature for the query image to the reference image. The GenerateIF function can take a choice variable which is 1 for reference image, and 2 for query image feature vector generation. Unfortunately, it's difficult to work out how to enforce the correct generation procedure from an abstract standing, where all features are effectively generic.
• 22.4.02 Webpages/Code: Updated the ID's page to include to the QbF algorithm IDs.
• 22.4.02 Code: Received the latest query by fax code to be integrated into the API.
• 22.4.02 Web pages: Moved last week's news to the news page

Week Beginning 15.04.02

• 19.4.02 Webpages: Updated the help pages, so they now include an example query by fax, and updated some of the words. Also packaged the help pages for those modules which IT-Innovations have, and sent the package to IT-Innovations, at 4:50pm.
• 19.4.02 Code: Still messing around with the Makefiles to try and get Mike's shared library stuff to work. It's now managed to mess most stuff up, so I'm not even able to compile Multiscale to test the new Multiscale PWT feature. Libtool is supposed to sort it all out for you, as far as I'm aware, but it doesn't seem to be.
• 18.4.02 Code: I've added a new function to the API which enforces the use of separate functions for loading header information from the data information within the feature vector files. This allows meta-features, like the multiscale code, to differentiate from data and header information. The compression routine in the Multiscale class then only compresses data. The matching will then only match on data. However, the matching on compressed data may only be viable for certain types of features. I haven't decided yet who should decide whether matching is viable. The feature module writer, or the application writer?
• 17.4.02 Code: Spent today attempting to get the library files working. There are static members in one or two of the classes which aren't too happy about being in shared libraries. Unfortunately, I've failed so far in sorting out the problem.
• 16.4.02 Webpages: I've updated the PWT, Query by fax, Stretcher Detector, and Stretcher Similarity help pages.
• 16.4.02 Code: I've now been given Mike's library code and a newer version of his MNS code to put into our newer system, which required a little bit of alteration. I've had to update the makefiles to recurse subdirectories and produce a .so shared library file for the cbir system.
• 15.4.02 Webpages: Updated the help pages of the multiscale variants to include a short description of the multiscale algorithm.
• 15.4.02 Webpages: Updated colour picker help page to include technical in formation and example results.
• 15.4.02 Code: Found a slight problem with the colour picker interface, that when a colour was selected from the colour palette, the colour patch was not being updated. This was fixed and a new package 5 was sent to IT-Innovations.
• 15.4.02 Web pages: Updated and moved the previous news to the news page

Week Beginning 08.04.02

• 12.4.02 Code: Checked the code for the colour picker interface, and after updating, packaged it up into package 5 which is read for IT-Innovations, which includes the code and classes, and default palette file and Lab colour patches.
• 12.4.02 Code: Updated the colour picker to load in pre-generated colour patches, rather than generating them on start-up. This makes it quicker. A new helper app, PatchGenerator, allows generation of the patches to JPEG files.
• 12.4.02 General: Sent package 4 to IT-Innovations.
• 11.4.02 Code: package4 is ready for IT-Innovations, which includes the Colour Matcher, the ColourHistogram, and the LabColourHistogram. This is the first code to be sent to IT-Innovations using the new based classes.
• 10.4.02 Code: Been completing the colour matcher code and removing memory leaks. Hopefully there will be no memory leaks in that code now (that includes the ColourHistogram and LabColourHistogram).
• 8.4.02 Code: Been updating the generic multiscale feature to conform to the Artiste API. I originally wrote it away from the API, so that I wasn't bogged down in some of the nasty code practices required while programming the API. I've now included calls back to the code I wrote from API compatible functions so that the Multiscale class is now FeatureVector compliant. However, to make it totally compliant I would have to implement an ImageProcessor for Multiscale objects, or multiply inherit from both FeatureVector and ImageProcessor. I'm not sure which is the best yet. Whichever, there will need to be special tests made in the generation/loading stage so that it will work with any type of vector. The create() method helps with this, but it does not get around all problems. I am also considering adding the predicate method isMultiscale() to modules so that the special case knows when to be run.
• 8.4.02 Web pages: Moved the previous news to the news page

Week Beginning 01.04.02

• 04.04.02 General: Came in to sort out the machine blanco, which contained the disc from the C2RMF containing the X-ray images which Fazly needs to continue his work on Craquleur detection and classification. However, blanco was not in an operational state, and needed reinstalling. This took from 9am to 12pm, and the subsequent copying of the 80Gb of data took from 12pm to 3pm. IT Innovations received the disc for re-use at about 3:30pm.
• 01.04.02 General: I'm on holiday (well, sort of) this week.

March 2002

Week Beginning 25.03.02

• 25.3.02 Code: Been continuing work with the generic multiscale implementation. So far it can decompose an image into a tiled pyramid and generate features for each of the tiles, and then save this out to a file. Next, the loader and comparer need to be implemented. There may also be the need to implement a compressor in a similar manner to Stephen Chan's Multiscale code.
• 25.3.02 Web pages: Moved last weeks's news to the news page

Week Beginning 18.03.02

• 22.3.02 Code: Been working on a generic Multiscale algorithm that will work with any feature vector type, in preparation for making a multiscale PWT. It will only be available in the newer code structure because the old image processing code structure does not support that type of functionality. The FeatureVector class now allows any feature vector to be accessed generically, and therefore allows a generic multiscale algorithm. This will also help later for a generic grid based algorithm for spatial colour matching.
• 22.3.02 Web pages: Updated the colour picker pages to show how to use the interface. The interface is a link from the colour picker module pages. Also added more navigation links back and forth between heirarchies.
• 21.3.02 Web pages: Added the colour picker help page (currently rather bare).
• 21.3.02 Code: Been testing the Lab Colour histogram, at various histogram sizes. For colour picking, the 10x10x10 histogram seems rather too precise, so I will try it at 6x6x6 next. Updated the new module code with the LabColourHistogram and made sure the stretcher detector code was working ok in it.
• 20.3.02 Code: Created an Lab Colour Histogram. Pretty easy from the new histogram code. Now generating 10x10x10 Lab colour histograms for our dataset.
• 19.3.02 Code: Been going through the code updating and making it more user-friendly. In particular I've altered the makefiles in the API classes directory and the Multiscale directory (the new directory). The makefile is much easier to use now, and should automatically detect dependancies, and should only require very small modification when a new module is inserted into the system.
• 19.3.02 General: I've moved the standard set of images that we use for the Multiscale demos into /data/archive/ppm. From there the directories can be linked to local directories for running Multiscale. This saves on disk space. There are scripts in the directory for automatically setting up the links for you.
• 19.3.02 Code: Updated the colour picker so that rather than buffering Lab->RGB value conversions, it buffers image slices in the RGB space. This makes the whole thing more responsive, and no more of that slider lag.
• 18.3.02 Web pages: Updates to the help pages include: technical description section on CCV, HistogramRGB and HistogramMono sections, as well as a navigation bar at the top, and a scale description and cross reference page accessed by clicking on the value of the module accuracy and module speed. Technical descriptions describe the feature vector, even with pictures. woo.
• 18.3.02 Web pages: Updated and moved last weeks's news to the news page

Week Beginning 11.03.02

• 15.3.02 Code: Been working on the latest colour picker interface. It now has an Lab colour patch from which colours can be selcted. The sliders are selectable between RGB, HSV and Lab, and the changes in colour are shown in a colour box, and in the Lab colour patch. Colours can be added to a set of colours to do a query on, and the various amounts of each colours can be altered, where the total amount of selected colour is not above 100%. As many colours as the user sees fit can be added to the selected colours and they can also be removed at will. You can see the colour picker interface (if you're Artiste internal) at http://corot.ecs.soton.ac.uk/~dpd98r/lab_picker.
• 12.3.02 Code: Made a new histogram viewer (just a simple one) using the xforms library, which loads in the new ColourHistogram ifo files and views them in a similar manner to the histogram shown on the right for monochrome images.
• 12.3.02 Code: Also managed to get the new Colour picker working with this new colour histogra m (it worked pretty much straight away, and works well).
• 12.3.02 Code: Completed the new implementation of the Colour Histogram module and got it working within both the old and new API code.
• 11.3.02 Code: I spent the day today re-writing Stephen Chan's HistogramRGB code so that it's more flexible, and generally easier to use and read. The new module will be ColourHistogram and it will make HistogramRGB redundant. (Although until there's an MColourHistogram, you will still have to use MHistogramRGB). This is in response to some requests at the user training meeting last week, however many of these will not be possible to be completed. For example, it was requested that the CCV be configurable also, but this will prove very difficult, again due to Steve's code. In particular, changing an underlying algorithm will not affect the multi-scale version, which proves even more of a problem, again due to lack of design in the object oriented nature of the code. Hopefully the new ColourHistogram module will also make the colour picker a possibility now, whereas before I was fighting against the histogram code.
• 11.3.02 Web pages: Updated and moved last weeks's news to the news page

• 08.3.02 General: Spent a very useful day with Christian who has given some good advice for the help pages, and also given some useful information regarding algorithm developement on cracks and stretchers and planks. Lets hope it's not all too late.
• 07.3.02 General: User meeting in IT Innovations proved useful again, with Marco, Christian, Gert, and Margret.
• 06.3.02 PMT Meeting: The PMT meeting in the V&A Museum in London seemed to go quite well. All involved were fairly happy with what was said, and the progress of Artiste is on schedule.
• 04.3.02 Web pages: Moved last weeks's news to the news pa ge

February 2002

Week Beginning 25.02.02

• 29.2.02 Code: Working on the multi-image histogram routine for plotting points. The image on the right is a grab of the dataplotter. Click it for a larger view. Happy St.David's Day.
• 28.2.02 General: Been re-installing RedHat on my system here, so not got any Artiste done. However, Miles now has the UV spot report up on his own site where he is able to maintain it. http://www.ecs.soton.ac.uk/~me k/UVSpotsDocs/index.html
• 27.2.02 Code: Updated the dataplotter, er, lots.
• 26.2.02 Web pages: Updated the help pages.
• 25.2.02 Code: Quickly knocked up a 3D data plotter with Java3D so that we may visualise the feature vectors that we are generating. This may allow us to determine whether clustering algorithms would work. The image on the right shows a screen shot of the data plotter.
• 25.2.02 Web pages: Moved last weeks's news to the news page

Week Beginning 18.02.02

• 21.2.02 Code: Changed an old image viewer, and made it work properly (!) so that it can be used as a sub-image selection tool for Artiste. I've sent this to Mike B. The online version can be seen internally at: http://corot.ecs.soton.ac.uk/~dpd98r/image_selector.
• 21.2.02 Code: Sent the first version of the colour picker to Mike, after the changes were made.
• 20.2.02 General: Faizal and I have just returned from a useful trip to see Mike Boniface at IT Innovations, about the usefulness of the PWT. It looks as though the PWT for fax retrieval is not very useful, as the correct results can be ranked quite low. Also Mike seems happy with the colour picker, so I'll just make some small changes then send that on.
• 19.2.02 Code: Mike's put out his first release to the team of his fvx replacement program, which should be much easier to use and more robust. The features are stored in an SQL database, rather than files, which also decreases the match time. I will attempt to plug this into the online matcher at some point, but maybe not soon. The readme for Mike's interface is http://corot.ecs.soton.ac.uk/~mrlw00r/cbir/readme.html
• 19.2.02 Web Pages: Found the problem with the colour picker, and that was that Microsoft (well, who else would it be?) are stupid, basically, and haven't implemented the Java interface correctly, causing the BorderLayout to have an incorrect default functionality. I've sorted it out now so the code doesn't expect default functionality and it works. I've also changed the classes so that they can be run as an application or an applet. If it's run as an application you can update the colour palette. If it's run in a web browser, this contravenes the security and so cannot be done. This functionality is therefore hidden when running as an applet.
• 18.2.02 Web Pages: Although not explicitly for the webpages, I've put a colour picker Java applet, which although simple is nice an effective. At the moment, I'm having some problems getting it running in both IE and Netscape. Here it is: http://corot.ecs.soton.ac.uk/~dpd98r/colour_picker ( internal only).
• 18.2.02 General: A trip to the IT Innovations Centre in Chilworth to meet with James and Rachael from the V&A proved very useful, certainly showing some of the limitations of both the Artiste interface, and the algorithms.
• 18.2.02 Web pages: Moved last weeks's news to the news page

Week Beginning 11.02.02

• 15.2.02 General: The user training is occuring on Monday morning, so it looks like we may be going up to.. hmm.. supervise, I suppose. Kirk, Miles and I also discussed colour and it looks like we may implement an Lab histogram and try some feature vector visualisations.
• 14.2.02 General: Happy Valentine's day.
• 13.2.02 General:Faizal and I are going upto Chilworth on Wednedsay 20th to IT Innovations to see about the PWT retrieval - and whether it's working correctly in the Artiste system. Hopefully we'll have a colour picker to take too.
• 11.2.02 General: John C's come to the rescue with a bug that Mike was having with his generic testing code and it's all been sorted. The problem was all to do with the way VIPS was being used to convert a 1-band image into a 3-band image.
• 11.2.02 Web pages: Miles has produced a short report on why the U.V. Spot detection measure is hard to implement. It can be found here: http://www.ecs.soton.ac.uk/~dpd98r/artiste/reports/UVSpot/1/UVWriteup.html.
• 11.2.02 Code: I've started development on a simple "find colour" matcher - which, yes, is what the HistogramRGB does, but this matcher will be geared towards finding single colours. It will use the HistogramRGB feature vector, to start with at least, but match only one of the bins - the bin containing the colour to be found. It should be stupendously quick for this reason.
• 11.2.02 General: The testing is still continuing. I left it over the weekend, after fixing some bugs in the sorting code (!) and sorting out the problems with regard to the integer overflow problems I was getting. However, over the weekend the thing crashed. It seemed to change directory and lose sight of the images, and therefore couldn't produce feature vectors. It completed the blur phase though. Now it's been started doing the rest, again.
• 11.2.02 Web pages: Moved last weeks's news to the news page

Week Beginning 04.02.02

• 04.2.02 Web pages: As a test of score normalisation I've swapped around the HistogramRGB and HistogramMono results after normalising their distances to between 0 and 1 on the online fvx, so that they now show similarity percentages rather than distances. It's only implemented for these two as these are the only two that we are sure of the maximum distance (which is 2).
• 04.2.02 General: I've generated the icons for the louvre set again into /database/icons/louvre.
• 04.2.02 General: Testing is still continuing as it takes ages on the Multiscale algorithms!
• 04.2.02 Code: Finally tracked down the problem causing the testing of the MHistogramMono to core dump. It was just trying to extract a test sub-image which was outside of the original image. This has been fixed and the tests are running.
• 04.2.02 Web pages: Moved last weeks's news to the news page

January 2002

Week Beginning 28.01.02

• 01.2.02 Code: Following Mike's discovery, I've implemented a function called GetRGBImage( VImage ) into the ImageProcessor abstract class (therefore available to all processing modules), which will convert LAB, and LABQ images to sRGB images, and also convert non-3-band images to 3-band. Other image conversions (e.g. from UCS, XYZ, etc.) could easily be incoorporated, but I didn't see the need right now.
• 01.2.02 Code: Pinch, punch, first of the month. Mike, while testing his test harness (!) found out that JPGs can in fact be 1 band, which is something we didn't know. This brought back into sharp relief the problem that CCVs and RGB and other colour-based modules don't work (i.e. crash!) on 1 band images. This will be fixed imminently. Good bit of detective work, Mike.
• 29.1.02 General: Update to below: I just realised they're using different matching code, although they should do the same thing, so there's still a problem. Been trying to fix it all day.
• 28.1.02 General: I've been continuing with the test process and have come across an interesting problem, which may be what IT-Innovations are experiencing. The MHistogramMono has been shown to work in a graphical test harness we have here, yet, seems to be ranking the correct match at 800+ in the batch-testing harness. They're using exactly the same feature vectors and the same matching code. What could be wrong?
• 28.1.02 Web pages: Moved last weeks's news to the news page

Week Beginning 21.01.02

• 25.1.02 General: To get some sort of quantitative measure for the border finder and classifier we need to know whether the classification of the borders by the neural network is correct or not. To achieve this I've written a little script that shows an image and allows input of a classification. The list is stored to a file which can be then used to work out stats. The script makes is easier and quicker to achieve this.
• 25.1.02 General: To start the day off I've started the resize tests on the HistogramMono module, now that the noisy tests are complete. Interestingly the noisy tests took much longer (6x longer) than the blur tests. This maybe to do with the way the populating of this histogram took place (it will be less sparsely distributed with noise). The resize test seems to be running as quick as the blur tests.
• 24.1.02 General: Finishing the day off with a whole bunch of test runs on the HistogramMono module. They're taking a while, as predicted, but I'll leave them going overnight. Testing under low pass filtering conditions is complete, and is now proceeding with testing under noisy conditions.
• 24.1.02 Web Pages: Made sure the basic help pages were complete, with the links to the border finder tests and extra information. The help pages are available at http://www.ecs.soton.ac.uk/~dpd98r/artiste/help.
• 24.1.02 General: I've altered Stephen Chan's test script to be a little more generic (or, at least, easier to run on various modules), and have run some tests on the monochrome histogram. Not sure what to do with the results, yet.
• 24.1.02 General: I've got a page up which shows some tests of the border finder on the Uffizi images on corot: http://corot.ecs.soton.ac.uk/~dpd98r/border/. It shows that the border finder is working correctly, and the neural net works fairly well, only mis-classifying 4 of the 14 paintings as circles, which are squares/rectangles. It's possible to understand how on one of the images, but rather unexpected for the others.
• 23.1.02 General: I've started running some tests to follow up my initial investigation on the border finder problems. I also intend to run the classifier over the whole Uffizi database to determine if there's a problem with the code (in general), or with the installation at IT Innovations. Initial tests who it is likely to be the latter, but by running the same test with our version of the code, we can determine when/where the error occured.
• 22.1.02 Code: Faizel has been looking at the changes to the PWT code, and has yet has not determined whether the changes have made any alterations to the expected output.
• 22.1.02 Web pages: Moved last weeks's news to the news page

Week Beginning 14.01.02

• 17.1.02 Code: Following yesterday's meeting with IT-Innovations, Miles and David looked at the reason the Border Finder and Classifier was not working on the Uffizi images. It turns out that the border finder algorithm is working well, and on our neural network the borders are even being classified correctly. This implies that there is something amiss with the IT-Innovations neural network. Investigation is continuing.
• 17.1.02 General: Miles now has an email address and can be contacted at mek@ecs.soton.ac.uk.
• 15.1.02 Web Pages: Added an image type input to the online FVX code which allows one to search based on the image type. For example "U.V." will find all the UV pictures, and "I.R." will find all the infrared pictures. Also added a "List Filenames Only", which will limit the database search to showing a list of filenames, which can be saved and used elsewhere, rather than the brief summary reports.
• 14.1.02 Code: Investigated the seemingly anomalous size of the monochrome histogram when used inside the multiscalar code. As it turns out, there's nothing at all wrong with the code, and the size of the feature vector for multiscale monochrome histograms is correct. The reason it's generally larger than the 3 dimensional RGB multiscalar histograms, is that the histograms are generally less sparsely populated than the colour counterparts, and therefore do not get compressed as effectively. One way to overcome the problem would be to reduce the number of bins in the monochrome histogram.
• 14.1.02 Web pages: Moved last weeks's news to the news page

Week Beginning 07.01.02

• 7.1.02 General: Welcoming Miles Kluth to the IAM Artiste developement team. Miles will be an employed RA on the team until the end of the project in July. He doesn't yet have a new user id, so has not yet got an email address.
• 7.1.02 Web pages: Moved last weeks's news to the news page

December 2001

Week Beginning 31.12.01

• 2.1.02 General: HAPPY NEW YEAAAAAAR!
• 2.1.02 Web pages: Moved last year's news to the news page :o)

Week Beginning 10.12.01

• 10.12.01 Web pages: Moved last week's news to the news page.

Week Beginning 3.12.01

• 6.12.01 Code: The plank detector is now into Mutliscale and works ok. It allows matching based on the number of horizontal and vertical lines it finds within the image. I've run it over the V&A directory too, just for fun, although the matching results aren't hugely useful.
• 5.12.01 Code: The plank detector seems to be running ok in fvx within the demo directory. I'm now just adding it to the Multiscale code.
• 4.12.01 Code: The plank detection (without the texture matching) has been received from Fazly. There is a problem with fvx with this code (on Fazly's machine), so I will try and integrate it into the old API code (in the demo), and once that is working roll it into the new code.
• 3.12.01 General: The fest is over, but nothing is complete yet.
• 3.12.01 Web pages: Moved last week's news to the news page.

November 2001

Week Beginning 26.11.01

• 27.10.01 General: It's fest time in IAM! The IAM fest, which is being held today, tomorrow, and Thursday, allows the group to work together on a number of projects, which will hopefully be finished by the end of the week. The Multimedia Understanding fest group, in which Kirk, Paul, Mike, Fazly, Faizel, and a number of other people are, will work on a similar system to Artiste, but from the point of view of a hand-held portable device in museums. The idea being that from the portable device users can find out information about a painting by pointing the device's camera at a painting. Content retrieval finds which painting, and retrieves information based on the result of the CBR and some context of the device's use.
• 26.11.01 Web pages: Moved last week's news to the news page.

Week Beginning 19.11.01

• 22.11.01 Code: The code for MNS has been completed, and is now being integrated into the old version of the test harness code. The reason for this is that this is the code that IT Innovations are currently using. Once this has been completed, it will be "upgraded" into the new code. IT Innovations will receive a copy of the MNS code that will fit into the older version of the API. It seems to fit into the old code quite well at the moment.
• 21.11.01 Web pages: I've added an upload-image querying ability to the online fvx test-harness. You can see it here: http://corot.ecs.soton.ac.uk/~dpd98r/test_db/
• 19.11.01 General: I've written some scripts (more robust than Kirk's) to go through a directory and all it's subdirectories and create thumbnails, or create feature vectors (using fvx). I'm running the feature vector generation script (with CCVs) on the dataset so that I can continue with the PHP interface to fvx.
• 19.11.01 Demo: Added Zepler and the Hay Wain to the demo along with some images that Kirk had taken from a video source of those images. They don't seem to work too well, though, so for this type of retrieval some pre-processing, or a different retrieval method might be required.
• 19.11.01 Web pages: Moved last week's news to the news page.

Week Beginning 12.11.01

• 14.11.01 Web Pages: For those with internal access, you might want to look at the database querying being done. It's basically a simpler version of the main database, but it should hopefully have support for content based functions via fvx at some point. Check out http://corot.ecs.soton.ac.uk/~dpd98r/test_db/.
• 13.11.01 General: Playing around with a PHP interface to fvx at the moment.
• 12.11.01 General/Code: Been getting a demo ready for Kirk to use in the Digital Imaging seminar tomorrow in the library. Using the old version of Multiscale and a special set of images from the V&A and the louvre.
• 12.11.01 Web pages: Moved last week's news to the news page.

Week Beginning 5.11.01

• 9.11.01 General: It has been decided that for the project review in December there needs to be a test version of [Q]MNS available to show, and possibly of Fazly's support detection algorithm. The V&A image set, once it's been downloaded, will be used to test some of the other algorithms, for monochrome subimage matching testing, for example. In the future, it is foreseen that a generalised Hough transform (GHT) could be used for butterly detection (and maybe internal shape feature and detail detection). Style classification may be achievable using a set of features matched against a known style, to find other similar images. It is also seen that some available texture measurement code could be used to test the PWTretrieval code against.
• 6.11.01 Web pages: Moved last week's news to the news page.

October 2001

Week Beginning 29.10.01

• 30.10.01 Images: The V&A dataset which is currently held at IT innovations is going to be retrieved to the University's servers, so that we can use those as a dataset... all 46,000 of them!!
• 29.10.01 Web pages: Moved last week's news to the news page.

Week Beginning 22.10.01

• 24.10.01 Code: Damn bugs.
• 23.10.01 General: There seem to be problems with corot at the moment, so nothing much can be done. Bug tracing is still happening when the machine returns.
• 22.10.01 Code: All the feature modules have been updated, however a bug has been spotted in the Multiscale program that is causing all multiscale-modules to fail comparison. This may not be due to the most recent changes. The bug is under investigation.
• 22.10.01 Web Pages: Due to the change in the ImageProcessor, the similarity functions no longer need to be known by the feature module implementer (other than StoreSimilarity), and so have been removed from the pages about the code structure.
• 22.10.01 Code: The code for similarity storage and retrieval has now been extrapolated from the HistogramMono class into the ImageProcessor class and compiles and executes correctly. Work will now go into removing code from the other image processors.
• 22.10.01 Code: There is still some code in the feature classes which is duplicated in each class. This can be overcome by moving these functions into the base classes. This will make the process of writing a feature module very easy. In the feature module file there will be just the functions that need to be coded: LoadIF, SaveIF, and in the image processor, GenerateIF, and CompareIF. Work is proceeding on achieving this.
• 22.10.01 Web pages: Moved last week's news to the news page.

Week Beginning 15.10.01

• 19.10.01 Web pages: Added the information for the image processing classes to the code structure pages which are rapidly becoming a quick reference guide to the API.
• 19.10.01 Trivia: The feature modules have reduced in size by about 300 lines of code due to the extraction of the memory blob I/O code. The PWTretrieval code has gone from 1135 lines to 817.
• 19.10.01 Web pages: Altered the steps to making a feature vector now that the base classes are available. If you're really interested, the old way of doing it is in an HTML comment in the code.
• 19.10.01 Web pages: Added the ability to download the library ready built for application development without the need to download and compile the library. Also added the latest image processing classes tar blob with the MemoryBlobWriter and the abstract classes.
• 19.10.01 Code: Updating of the feature modules to use the MemoryBlobWriter centralised code is complete.
• 19.10.01 Code: A new class, MemoryBlobWriter, has been defined which contains all the implementation for the reading and writing to memory in static functions. This avoids the need to clutter the feature modules with this, duplicating the code. The current feature modules are being altered to use this class rather that their local versions of Read and Write.
• 19.10.01 Code: As well as ensuring that the modules are still API compliant and work correctly with the new abstract base classes, work has begun on moving all the memory blob and file reading and writing code out to a separate (possibly statically callable) class. This should clean up the modules alotand again reduces the amount of code duplication.
• 19.10.01 Web pages: Added more to the code structure pages to include a very brief overview of the API functions.
• 18.10.01 Trivia: The Multiscale harness has halved in size due to the abstract base classes: 2600 lines of code to 1300.
• 18.10.01 Code: Updating of Multiscale is complete!
• 18.10.01 Code: Some extra work was needed to make sure all the feature modules correctly inherited and conformed to the stricter API. Work is now proceeding in testing that the feature modules can be used in the way intended, without too much code re-writing. This is achieved by altering Multiscale to use the new class heirarchy.
• 17.10.01 Code: The updating of the feature vector and image processor definitions have been completed and they now inherit from the abstract base classes FeatureVector and ImageProcessor, although fully testing this has not yet been achieved, and therefore the code tar-blobs below remain in the previous state.
• 17.10.01 Web pages: Added a code structure page.
• 17.10.01 Code: Work is continuing on converting the old feature modules to the improved OO design. Many bad coding problems have been discovered which are being fixed in the meantime, but it slowing down progress. The multiscale features all referred to each other's private members. This is pretty much fixed and encapsulated now except for one call from the multiscale code to the non-multiscale GenerateIFAux private member. This means the multiscale class has to be a friend to the non-multiscale class. A bit nasty. That will be fixed in due course.
• 17.10.01 Web pages: Added the news archive section to keep the index page smaller. The news section will now contain a date and type of news.
• 16.10.01 Code: The classes ImageProcessor and FeatureVector have been defined which set out the abstract methods which an image processor or feature vector need to comply to. This has been tested to comply with the API by inhereting the HistogramMono class from these and compiling successfully.
• 15.10.01 Code: Work is going ahead on making the API code of a better object oriented paradaigm (as stated in the "stuff to do" pages). The code for the feature vectors and image processors is based very much on a procedural programming paradaigm, despite being implemented in an object oriented language. This makes for a lot of wasted opportunities. The use of the API should not change at all, unless the application programmer wished to use the extra advantages that the OO paradigm facilitates. An example of the advantage is having a MultiscaleFeature class that provides the multiscale ability on any feature without the need to create new features like MCCV, MHistogramRGB, etc.

Old News

• 12.10.01 Web pages: Added a progress and contacts web page.
• 11.10.01 Code: All updates have been made, and importantly the .c files have all been renamed to .cpp for the convenience of IT Innovations, and it's important that all new modules should be named in this convention. The tar blob below have been updated. The updates have also been rolled into the templates and the new modules.
• Received a package of changes made to code by IT Innovations. These changes are currently being implemented into the base set of code. Once this is complete the changes will be filtered into the new feature modules.
• Updated the Multiscale program to deal with command-line parameters, with two new ones to allow for varying the location of the reference or query images. This should allow us to use a centralised repository of feature vectors for testing, without the need to install the test images, etc. Also the names of the default directory for the reference images and the file containing the filenames have been altered from CCVedImages to ReferenceImages, and from CCVDatabase.db to ReferenceDatabase.db, respectively. With the new switches -r and -d you can set the directories in which Multiscale will work. The tar blobs for Multiscale, the test images, and tools have been updated below.
• Completed the updating of the ifos which are part of the test set of images available as a large tar blob below.
• Integration of the PWTretrieval into Multiscale and fvx is complete! All the code tar blobs below have been updated.
• Completed (for the moment) the restructuring of the directory, er, structure.
• The pyramid structure wavelet transform (PWT) retrieval module is complete!! It's currently being integrated into the latest fvx.c and Multiscale.c to keep the code upto date, although it is effectively ready for IT innovations. Super work Faizal.
• The gray level histogram (GLH) retrieval module is complete!! It's already been included into the latest version of the fvx.c and Multiscale.c and is ready for IT Innovations. It will be tested alongside the latest retrieval modules. A single-GLH to multiscale-GLH (MGLH) module is underway.
• The hough transform based support detection module is also underway by Fazly.
• The multi-modal neighbourhood signature (MNS) is near completion to be integrated with the latest and greatest versions of fvx.c and Multiscale.c.