Would you be interested in the future to do custom development?
Definitely. Specially during these first releases. The API is getting stable and as usual lots of feedback on how things should be is needed from users.
How can I be notified of any new information regarding the future of this project ?
All news and improvements are posted at http://server.imgseek.net which can be subscribed on an RSS reader or by email. User and development issues are discussed on the official mailing lists. Also please don’t hesitate in asking for more details if anything is not clear.
How about an approach where we can color index images of our used cars (say we have images of the cars, but not the colors of the cars). We would like our users to be able to find cars in similar colours. Do you think that isk-daemon can be used for indexing ~50.000 cars photos in 10-20 major color categories (Black, grey, brown, green ….)?
What happens is that for such a technology to give you exceptional results it would need to actually recognize or get a reasonably good approximation of what is the car and what is not when looking at an usual car photo with some other background and image features with different colors that contrast with the actual car color.
The application programming interface (API) for interacting to isk-daemon doesn’t currently suppport this type of query but it is on the roadmap for the next months. Please leave an email or subscribe to the RSS feed to keep updated on the progress.
Will isk-daemon find ’similar’ images and pass back a ‘list’ of image files (filenames) ?
The isk-daemon image server would return a list of image IDs and not filenames. These image IDs (long unsigned integers) are the same you supplied to the isk-daemon when you added the images to its database.
Can I get an hash or array from isk-daemon?
Yes, you can. The result for a similarity query is an XML describing the similar images or an array represented using the RPC technology you choose (SOAP, XML-RPC), so you can transform it into an array or similar data structure that is native to your web frontend code.
What are the advantages over GIFT?
We believe the main advantage over GIFT is the query results quality, speed (isk-daemon has a very optimized C++ implementation with smart query algorithms based on Wavelets) that takes a very different and simpler approach that GIFT takes and finally, the simplicity of integrating it into your existing web application.
How accurate is it? How can it be made more accurate and how easily?
It’s hard to measure it’s accuracy, but results are quite good. We should test it on a real database and system to see if it’s good enough for your needs.
Does it get more accurate with the more images you put into a database?
Yes.
Do large image databases make it slow down?
See the Scalability section.
What is the order of image matching within the multiresolution wavelet decomposition (colour, shape…)?
All these factors are taken into account at the same time, but the importance of shape versus the importance of colour could be fine adjusted. This option is to be added to the API on the next versions.
Is this matching information code compatible with other software or could it be made to be so?
The matching information code (that is, an image fingerprint once a new image added to the database is processed and stored on the image similarity database) is not compatible with other systems, due to isk-deamon unique image similarity algorithm.
How can I improve content-matching for the typical images I’m working with?
There is no simple answer. These are some of the things you could do to provide better matches when considering your specific problem:
1) tweak the internal weights used on the wavelet haar decomposition and index/signature building
2) pre-process your target and training images in order to remove background noise: use face detection and image segmentation techniques to separate what is a face from what is not and only supply to isk-daemon a pre-processed version of images where only faces appear
Regarding the first approach, the weights/parameters you mostly would need to change are
const float weights[2][6][3]
which can be found at https://imgseek.svn.sourceforge.net/svnroot/imgseek/net.imgseek.imgdb/trunk/src/net/imgseek/imgdb/core/imgdb.h
You may try running experiments against a control group of photos to see which variations on these weights cause a smaller rate of classification errors.
To do so you would need to do some refactoring in order to expose these weights through the imgdb.h API so they can be varied dynamically from your calling code.
Source code for the image similarity processing algorithms can be found at https://imgseek.svn.sourceforge.net/svnroot/imgseek/net.imgseek.imgdb/trunk/src/net/imgseek/imgdb/core/
This when compiled generates dynamic libraries with bindings for Python and Java.
Where is the user interface showing thumbnails?
isk-daemon does not come with an end-user interface for querying for similar images. You will need some PHP coding skills in order to code a working web screen that should make a call to isk-daemon, understand the query reply a list of image ids and similarity ratios) and from the replied image ids, construct an html screen where the image thumbnails associated to these ids are shown.
Please take a look at the sample PHP code which queries isk-daemon for images similar to the one having the id “1″ and prints the resulting image ids on the web page. You would need to adapt it to generate the HTML code for an image gallery browser, with <img> tags with their “src” attribute pointing to the URL where the images associated to these IDs are being served to the end-user.
isk-daemon does not keep track of image filenames and the pixel data associated with a given image id. Your frontend web system should know how to generate a page containing thumbnails and the server path where the images associated to these IDs are.
isk-daemon was built this way in order to be flexible and to allow it to work integrated to any image-related web site. Most of them already have a database associating its internal image ids to image metadata like filenames and so on, so there is no point in duplicating this data inside isk-daemon.
How exactly does the similarity work? What about the wavelet and image processing technology behind imgSeek and isk-daemon?
All metric and query ideas involving 2D wavelet transforms were based on the paper Fast Multiresolution Image Querying by Charles E. Jacobs, Adam Finkelstein and David H. Salesin, which is available at http://grail.cs.washington.edu/projects/query/
The C++ code at imgdb (part of imgSeek and isk-daemon) is almost a direct implementation of the ideas on this paper.
Is it possible to access the representation through the database file or similar?
Image signatures (wavelet transform coefficients) and data structures for coefficient buckets are serialized to disk using an internal
binary format. You can see how they are serialized by looking at the savedb() C++ routine at the imgdb source module. (both at imgSeek source code or at
https://imgseek.svn.sourceforge.net/svnroot/imgseek/trunk/net.imgseek.imgdb)