Hello, my name is Peter Bell from FAU and I want to introduce you to images.ai.

Images.ai is a retrieval and explorer interface for images mainly developed by Fabian Offert

with some support by Oleg Halamov and me.

So first you log into the interface of course and you directly will see this beta report

bugs option as we are in the beta version.

We would be very happy if you can give us feedbacks to bugs here or as a feedback on

the hi at images.ai email.

Then we have settings here.

Here you can define how many results you want to have in every search.

I raise that to 50 and you can decide which size the images should have 128 pixels maybe

okay.

So I apply.

Of course you have a longer help section where you find everything.

So going back to the interface you will see the number of pictures of the data set here

and the data set name.

This is a Metropolitan Museum of Art.

We have four museums the Metropolitan, the MoMA in New York, the Harvard with a collection

of scientific instruments and the Rijksmuseum which was a part of its open access collection.

We have some particular data sets from annunciations, reception and celebrities.

Starting with the MET you can choose images directly from here by clicking and putting

them up or you can of course upload images from your desktop.

Let's try that.

This is landscape and you directly will get other landscapes.

You see different formats, different regions and styles.

So you may add other positive examples which you think fits to that or you can put images

on the negative side so that the computer trains them to not give results like that.

And of course you can remove images or flip the sides here and put them to the positive

side.

You already saw that every change occurs to another ranking of neighbors here and you

can change these results also with the embedding here.

We have VGG19 which is a convolutional neural network, so a deep learning approach, but

we have also the very simple raw comparison of the pixels.

So we now get results which have somehow the same texture or paper color but as you see

these raw pixels comparison only helps in some cases.

So when we change the embedding from raw here to poses we hope to find another pose comparable

to that dancer.

And we see at least there are other poses in the images but of course this very complex

pose isn't detected quite well.

So at that point we may not only rank the results of these different images here in

different results belonging to that or that image but we may have kind of a medium of

these three images in its pose with this centroid.

And you see in this first line that we really have some pose which is comparable to all

these three.

At last we could also change the distance and at the moment we have Manhattan, it can

be Euclidean or angular.

This only makes normally slightly changes but sometimes it's also interesting to try.

By that you already see the very basic functions of images.ai.

One very neat feature is that you can change between the different data sets and taking

images with you.

So you may go from Metropolitan to the Rijksmuseum.