AI functionality: My first impressions and findings using “AI4PSU”

Post Reply
bimo
Posts: 97
Joined: 29 Aug 07 22:34
Location: Germany

AI functionality: My first impressions and findings using “AI4PSU”

Post by bimo » 26 Sep 19 15:58

AI detect labels
  • I like the functionality and especially the possibility to differentiate between relevant labels (which have to be moved manually from the “AI Catalog Labels” to one’s own label hierarchy) and non-relevant labels which remain in the “AI Catalog Labels” section.
  • The auto-assign feature is a bit critical though: if Google Vision provides a wrong result (which of course might be a correct result for a different image and therefore might be an existing catalog label) it will be auto-assigned nevertheless:
    Taking e.g. an image with just one dog of breed “Sloughi” (an Arabian sighthound), Google Vision wrongly pretends to detect three breeds(!) for this dog: “Anatolischer Schäferhund”, “Deutsche Dogge” and “Whippet”. Of course only one breed can be the right solution for one dog, but moreover the right value (“Sloughi”) isn’t even part of the delivered results! But as “Whippet” (a little English sighthound) is a valid entry in my catalog, “Whippet” is assigned by mistake. If “Anatolischer Schäferhund”and “Deutsche Dogge” would have been entries in my catalog they would have been assigned as well...

    That’s why auto-assigning catalog labels seems to be questionable to me.
AI Object Recognition and AI Landmark Recognition both aim at the new category „Objects“. IPTC has a section “Artwork or Object” which aims at the same or at least similar content. As always in PSU you can rename the category to your like; I renamed it as “Artwork, Landmarks and Objects” and would want to subdivide the section into these parts as well. At least Landmarks and Objects will then be AI-driven while for the time being Artwork will be handcrafted.

AI Object Recognition
  • Each and every recognized object is written as and auto-assigned to a new catalog label (or added to an existing label) in section “Objects”. In contrast to “AI detect labels” no differentiation can be made between relevant and non-relevant catalog labels (objects). So either one has to accept lots of non-relevant labels or to clean the hierarchy each time an unwanted label appears.

    IMO a section “AI Object labels” would be helpful:
    • For each object delivered by Google Vision which can be found in the AI Objects Labels: forget it, because it is not relevant
    • For each object delivered by Google Vision which can be found in the regular objects’ catalog section: auto-assign the object
    • For each object delivered by Google Vision which can neither be found in the regular objects’ catalog section nor in the AI Objects Labels :
      • If it’s relevant: assign the object manually and write a new label to the regular objects’ catalog
      • If it’s not relevant: write a new label to the AI Objects Labels
  • Moreover ambiguities are still possible: the “Krusenstern”, the famous ~100m Russian sailing vessel, is without doubt a “ship” and not a “boat”, but Google Vision on the one hand pretends it to be a ship then again calls it a boat. Personally, I would try to solve this situation by creating a new label “Vessel” with synonyms “Ship” and “Boat” – but the algorithm behind AI Object Recognition doesn’t allow to find and assign a catalog label by use of synonyms.

    IMO the use of synonyms is mandatory for assigning the appropriate label.
  • If you try to structure the Objects Section according to your requirements, that will be in vain as the algorithm can’t handle this neither. So all you can expect is an alphabetically sorted list – quite confusing IMO bearing in mind the lots of labels.

    IMO the extension by manually added structures (= additional catalog labels) is important for the usability of the new category “Objects”.
    (Feature Request: https://mantis.idimager.com/view.php?id=3378)
AI Landmark Recognition
  • I’m really overwhelmed by the huge amount of landmarks being found by Google Vision. But as I’m not interested in e.g. Restaurants and Cafés, I would also want to differentiate relevant and non-relevant landmarks.

    IMO an equivalent section “AI Landmark Labels” (containing the non-relevant landmarks) would be helpful.
  • The name of a certain landmark is not necessarily identical across the images being analyzed.

    So again, synonyms are needed.
  • To structure the landmarks I would sort them according to the countries they can be found in.

    Again, this means that existing labels for landmarks must be found regardless of their hierarchical position.
    (Feature Request: https://mantis.idimager.com/view.php?id=3378)
  • As the landmark is what is been shown in the image, its GEO Location belongs to the IPTC structure “LocationShown” (corresponding GPS Info is to be stored in EXIF GPS (34853:[19-26], 0x8825:[19-26])). But if you choose to “Overwrite pre-existing GEO information” the position is wrongly written to EXIF GPS (34853:[1-6], 0x8825:[1-6]) which contain the GPS data for “LocationCreated”.

    IMO this is a fatal bug because it violates each and every appropriate standard! (Bug report: https://mantis.idimager.com/view.php?id=3376)
    (More details about LocationCreated and LocationShown may be found under https://mantis.idimager.com/view.php?id=2103).
Michael

peterpix
Posts: 39
Joined: 01 Jun 16 15:36

Re: AI functionality: My first impressions and findings using “AI4PSU”

Post by peterpix » 30 Sep 19 16:11

Michael, I assume that like me I had to sign up with Google to be billed for the pleasure of using their API? They graciously offered to give me £300, but I have pressed the AI Detect Labels button I think 4 times and my £300 credit is now £243.90. The labels that were returned were not breathtaking either. So far anyway it hasn't ID'd anything or anyone. For example a famous seaside cliff view returned Headland, Coastal and oceanic landforms, and Klippe (which is German for cliff). So so far I am not blown away by this AI feature, and certainly not by the considerable cost of each button press. Maybe I am using it in the wrong way? Would be interested in others' views and experiences.
Last edited by peterpix on 30 Sep 19 16:57, edited 1 time in total.

Hert
Posts: 5866
Joined: 13 Sep 03 7:24

Re: AI functionality: My first impressions and findings using “AI4PSU”

Post by Hert » 30 Sep 19 16:34

I believe you receive 300 USD with your Google signup. That probably converts to 243 GBP.
This is a User-to-User forum which means that users post questions here for other users.
Feature requests, change suggestions, or bugs can be logged in the ticketing system

peterpix
Posts: 39
Joined: 01 Jun 16 15:36

Re: AI functionality: My first impressions and findings using “AI4PSU”

Post by peterpix » 30 Sep 19 17:09

Ah, maybe that's it, thanks. Assumed it was dollars. Do you know the approximate rate of attrition per button press? Is it something so low it doesn't matter e.g. $0.001?

Hert
Posts: 5866
Joined: 13 Sep 03 7:24

Re: AI functionality: My first impressions and findings using “AI4PSU”

Post by Hert » 30 Sep 19 17:54

In the Help section (click the question mark in the left sidebar) open the Artificial Intelligence Quick Start manual.
https://cloud.google.com/vision/pricing
This is a User-to-User forum which means that users post questions here for other users.
Feature requests, change suggestions, or bugs can be logged in the ticketing system

bimo
Posts: 97
Joined: 29 Aug 07 22:34
Location: Germany

Re: AI functionality: My first impressions and findings using “AI4PSU”

Post by bimo » 30 Sep 19 22:21

Peter,

my current balance is 0,16€ versus a credit of about 260€. So, no worries...
Michael

Hert
Posts: 5866
Joined: 13 Sep 03 7:24

Re: AI functionality: My first impressions and findings using “AI4PSU”

Post by Hert » 03 Oct 19 8:43

I made this a sticky topic. Feedback about AI features can be collected here.
This is a User-to-User forum which means that users post questions here for other users.
Feature requests, change suggestions, or bugs can be logged in the ticketing system

Jesperben
Posts: 48
Joined: 02 Jul 11 8:07
Location: Odense, Denmark

Re: AI functionality: My first impressions and findings using “AI4PSU”

Post by Jesperben » 10 Oct 19 13:24

I was also a bit underwhelmed by the service. It didn't recognize any of welle known danish structures - the parliament, iconic bridges and others.

But I think this is a very important path, and I am sure the technology will be very helpful - but maybe we will have to wait a little for the big benefits.
Trying to get a lot of old pictures out to the family

Post Reply