IMatch 2020 - Overview Video with all major new features

[Mario]

Hello Users.

IMatch 2020 is almost finished. And it's awesome!
The tester groups locally and on the web are working hard to make this new version working smoothly. Thanks to all of them.

I have produced a video which shows the major new features in IMatch 2020. You will enjoy this.

IMatch 2020 - What's New?

Watch the video on YouTube and let me know what you think in the comments below.

[hluxem]

Great job Mario!

This looks really great. I hope we don't have to wait too long for all these new and exciting features. I specifically look forward to the new features with face recognition and working with people.

Thanks a lot for the video,

Heiner

[Tveloso]

Absolutely awesome!

I too am eagerly looking forward to the Face Recognition functionality...and the Auto Tagger also!!

The video mentions that the Auto Tagger is fully integrated with the Thesaurus...I look forward to seeing that as well!

Mario is just amazing!

[mastodon]

Fabulous! Face recog, handling people and events like this is huge step forward for family photos like me. The other small things are also very-very good ideas. One can clearly see, that your developments are driven by the needs of the customers (not by your market division - if you would have one).
GREAT work, thank you!

[Aubrey]

face recognition useful

[PaulS]

Wow   

[Darius1968]

What I got out of this video is looking forward to the new Events feature that seems like it should answer someone's prayers for whom there has been this want of a very user-friendly way of organizing photoshoots!  Just create the well established user-stack for an event (photoshoot), and then, IMatch has now broken new ground in offering the convenience and user-friendliness of a new dashboard to name and categorize these events! 

[Mario]

Events are just another great way to visualize what's in your database.

You could achieve something similar using an "Event" category hierarchy - including a thumbnail representing the event (the category tree can show thumbnails instead of icons) and color-codes to differentiate different types of events. Using a naming scheme or manually reordering the event categories allows you to bring the events in a specific (time line) order.

Actually, many users do that. Which is why I allow to create events from categories - as an easy way to migrate or reuse existing category-based schemata.
Other users create one folder per 'event'. Which is why events can be based on folders - again to facilitate reuse of what organization users already have.

The new IMatch Events are easier to understand for new users (I suppose and know from test users). And maybe more intuitive than categories for many.
Especially if you use the "Create event from these selected files" or "Create event from all files between start date and end date".

The timeline in the Event View is quite pretty and very 'visual'. Which should appeal especially to new IMatch users. But the old ones also like it 

One of the points that does not come across in the video is that many of the People/Event features are implemented as IMatch apps. The UI you see is actually a Chromium browser window embedded in an IMatch panel. All this is possible thanks to IMatch Anywhere WebServices running in IMatch.

Using modern web technology allows me to implement things faster. To change things faster. To utilize all the powerful visualization and animation technology available in modern web browsers. And IMatch 2020 is the first version of IMatch which truly integrates this technology into built-in features. I expect lots of other cool and useful enhancements based on this tech stack.

[Menace]

@Mario: So I can start to advertise IMatch 2020?

[Mario]

@Mario: So I can start to advertise IMatch 2020?

You can always advertise IMatch 
Word of mouth is the best advertisement - and the only I can afford.

But please keep in mind that there is no release date for IMatch 2020 yet.
With the video I've just wanted to share some info about what to expect and to show to users what I've been working on over the past 8 to 12 months.

IMatch 2020 is in RC state (Release Candidate) so I won't make any substantial changes anymore.
But I'm still finishing up some loose ends, documentation, help system, web site updates, final test rounds, updates for the automatic build pipeline and stuff like that.
Testers can find bugs which I need to fix before shipping etc.

As always, I rather spend a week more on the product than to ship too early.

[Menace]

Ok, than I wait a little bit longer. 

[Mario]

But all users of course should search their pockets, couches, car seats etc. for some spare change - to be able to afford the upgrade fee for IMatch 2020.
IMatch 2019 was free (I hope you all remember that) but IMatch 2020 is not a free update.

[Menace]

Already waiting for paying. 

[Bri Wit]

Looks awesome Mario! I look forward to it. A free upgrade to 2019 was a nice surprise. You have created an awesome program and provide stellar support. I will happily pay for the upgrade.

[jcldl]

Everything looks great. I am eager to see all these so many improvements!

[dcb]

Sent the video to a friend. Told him that's why I love IMatch and even gave him a figure of what I'm prepared to pay for the upgrade. 

No doubt I've got some Who and Event categories to think about differently.

[axel.hennig]

Looks really great.

I would be more interested in the following (topic "Face Recognition"):

Assume I've already got a category structure like:

People
  - Anna
  - Bernd
  - ...
  - Zoe

Would it be possible to tell the AI that I've got already some information (which people are on which photo), so that the training of the AI would be faster and possibly more "correct" in guessing which face belongs to which person?

[Mario]

I would be more interested in the following (topic "Face Recognition"):

Assume I've already got a category structure like:

People
  - Anna
  - Bernd
  - ...
  - Zoe

Would it be possible to tell the AI that I've got already some information (which people are on which photo), so that the training of the AI would be faster and possibly more "correct" in guessing which face belongs to which person?

No.
The AI detects faces and creates a 'fingerprint' of each face, compares it with fingerprints of the faces of known persons and then automatically links each face to a person.
You create persons up-front or on-demand. You assign each person to a couple of faces to train the AI.
The AI is then usually able to find all other images of that person automatically.

Keywords or categories assigned to a file are irrelevant for face detection or face recognition.
An image can have any number of keywords, can be in any number of categories and can have any number of faces.

Of course purpose-built solutions like "If a file is in one of these categories (selection) and only one (!) face is detected in the image, lookup the matching person by comparing the tag/full name/short name (?) with the category name like this ... and then assign this person".
This might work for your case. But if you ask 10 users you will get 12 different ways to do this, 12 different setups, hierarchy depths, organizations, special cases and exclusions for such scenarios....really not worth the trouble.

The automatic works great most of the time. You have to see it to believe it. Unless your files are very special, and massively different from the 100,000 files with faces test sets I've used for development, testing and verification.

Your existing organization may help you later, when you confirm the persons assigned by the AI.
In combination with the people filter you can confirm large sets of files in batch using your categories with a few mouse-clicks.
If you have several 100,000 files or hundreds of different people to work through, this will save you time.

[Tveloso]

Mario, what about existing Face Annotations?  Will the AI make use of those?...or will it be starting "from scratch" with Face Detection and Recognition, independent from the already marked faces?

Even if it's the latter, I imagine that it will be quick if the confirm process is done from within the Annotations Collection, as you describe above for Categories.  But I was just curious if one took the approach of creating the persons up front, if the existing Face Annotations could be "pulled in" to train the AI.

[Mario]

Existing plain face annotations are converted into real AI-backed face annotations during the initial migration process.

[Tveloso]

Excellent.  Thank you so much Mario!

[Mario]

You're welcome.

The existing face annotations are detected, converted and persons are created using the "tags" assigned to the existing annotations. This groups all files with the same annotations under the same "new" person. After the database migration is complete you can modify the created persons, change the tag, name etc. It's all pretty comfortable and automatic. May take a while when you have thousands of face annotations, but that's OK I guess.

[mastodon]

I am really curious. I have about 36.000 photos, and 12.000 of theam are more or less face tagged.
Will the AI find (and more or less recognise) all possible new (untagged) faces in one session? Or does it have a daily or any limit?

[Mario]

I am really curious. I have about 36.000 photos, and 12.000 of theam are more or less face tagged.
Will the AI find (and more or less recognise) all possible new (untagged) faces in one session? Or does it have a daily or any limit?

During the first open, IMatch converts your existing face annotations into 'real' IMatch 2020 face annotations.
IMatch uses the 'face rectangle' you have defined already to detect a face (if possible) and to create a face fingerprint for it.
It also creates one person for each unique tag you have used with your annotation.
After this step this has completed, the People View contains one person for each annotation tag, and the files for each person.
This has also trained the AI to recognize these persons.

You can now review the persons, add more data (like a full name, birth date etc.).
The birth date allows IMatch to display the age of persons relative to the image date. Which is pretty cool.

Then run face detection on the remaining 24,000 images in one batch or as it is convenient to you.
IMatch by default skips files with existing annotations when searching for faces. So, you can select all your 36,000 files and let it rip.

24,000 files is not much. On a fairly modern PC (say, 4 years old, 4+ cores) it takes probably about one hour to process all the files (detect faces and assign known persons). Notebooks are usually slower.

You can then use the People View and Filter to review the results, set / change person assignments as needed. It's a straightforward and fast process.

There is no daily limit or anything.
The AI for this is built into IMatch and runs on your PC. No external service is used, no data ever leaves your PC.

The performance of the AI is only limited by the speed and number of processors in your computer. The more, the better.

[mastodon]

That is much-much faster then Picasa.
Can the AI use the video card, my proc is just an I5 3570, but the video card is an MSI R9 270x Gaming 2GB?

[Mario]

That is much-much faster then Picasa.
Can the AI use the video card, my proc is just an I5 3570, but the video card is an MSI R9 270x Gaming 2GB?

No. GPUs with tensor or CUDA cores are helpful while training new AI models. Throwing two or four GPUs at the problem reduces the model training time to a week or two... at a cost. Or just rent the computing power needed from a cloud vendor.

On the end-user side, this is not needed. At least not for this domain.
It would also be a really bad idea for IMatch to require specific graphic cards or the installation of complex and somewhat instable nVidea driver stacks.
This would be a support nightmare. Shudder.

Even the simple requirement of processors with support for AVX (which is standard since 2011 !) has caught some users unaware - because they use very old or low-end PCs without AVX. An now cannot run IMatch 2020  without upgrading their hardware. This affects of course only a handful of users and they can switch to IMatch 2020 when their steam-driven computers bite the dust 

Typical face detection with standard settings is 0.5 to 1 seconds per image (most of the time is spent loading the image so the GPU would do nothing anyway).
And IMatch can do that in parallel on all cores in your system.

Running face detection on a 12 core + HT Radeon CPU is a sight, I can tell you. All cores maxed out, churning through thousands of images in almost no time.

But is also works just as well on 2 core Notebook CPU - just slower...

[ben]

This sounds soooo great. Can't wait to spend my money.

I have followed your older posts about face recognition. At that time a local AI was no option and you were trying cloud solutions. May I ask how it works in the background? What do you need the different cloud AI services as can be seen on the video?
Ben

[DigPeter]

Is the face recognition facility limited by the aspect or size if the image?  I do not have many "faces" amongst my photos - mainly family.  These are tagged with their names.  These "faces" are taken at all angles and distances.

[Mario]

This sounds soooo great. Can't wait to spend my money.
What do you need the different cloud AI services as can be seen on the video?
Ben

These are for AutoTagging (unrelated to face recognition).

The new IMatch AutoTagger allows you to automatically add keywords to your files. Using specialized 3rd party services.
I don't want to force my users to use a specific service so IMatch supports the four top AI tagging services available today: Clarifai, Google, imagga and Microsoft.

[Mario]

Is the face recognition facility limited by the aspect or size if the image?  I do not have many "faces" amongst my photos - mainly family.  These are tagged with their names.  These "faces" are taken at all angles and distances.

No.

[Winfried]

Does the face recognition work with b/w-photos or scans?

Winfried

[Mario]

Does the face recognition work with b/w-photos or scans?

Winfried

Yes. Color differences, age differences, different hair styles etc. are handled very well by the AI.

Face detection or recognition may fail if your images contain a lot of noise, badly blurred faces, very small faces, scratched faces. etc.
Scans from old historical images sometimes cause this kind of unique problems.
Hard to tell without a sample. You should have attached a JPEG sample. I could have tested it and show you the results.

[mastodon]

I have tagged not just faces, but heads from behind. Will this confuse the facerecog AI?

[Mario]

IMatch cannot identify faces from the back of the head 
Such annotations, when imported during the initial migration, will become true manual face annotations in IMatch 2020 terms.

These are face annotations created by the user which are not backed by an AI face fingerprint.
For example, if IMatch cannot detect a face because is partially obscured, in shadow etc.
Users can always create face annotations manually - even if IMatch cannot detect a face in that "box".

The AI cannot automatically assign a person to such a manual face annotation. That's the only difference.
You can still assign a person to such annotations and they will then work like any other face annotation (show up in the People View, automatically assign keywords and categories etc.).
This is all part of the People concept in IMatch 2020.

[altae]

Very interesting new features. If they work as described buying the upgrade will be a no-brainer to me.

[Mario]

The Beta testers seem to like it very much 

[kiwilink]

Can't wait to use the attribute changer to correct many of my pics with the wrong date.  Thanks for adding this Mario!!!

[Mario]

TimeWiz is the name