Google AI and PSU design

Post Reply
snowman1
Posts: 329
Joined: 01 Jan 07 3:13
Location: UK

Google AI and PSU design

Post by snowman1 » 21 Jun 21 19:50

I've posted before about Google AI being inaccurate. I think PSU's design in this area exacerbates the problem. It allows us to "detect" the labels and then decide whether to allocate them. So far so good. But it seems that if a label detected by G-AI pre-exists in the PSU catalog's main label hierarchy then it is allocated automatically. This to me makes things worse, because often the label G-AI has come back with is either wrong or unwanted, but now I have to click to un-allocate the unwanted labels from the image(s) if that image was in the main catalog.

As an example: G-AI often comes back with "animal" or "plant" on my landscape pictures where there are no animals and... well it's a landscape so it has millions of plants but it's not a useful label. But because both labels are in the main catalog (being useful in other contexts) they get allocated to the images. I can see the reasoning behind the design, but that reasoning is predicated on the labels always being accurate which is, in reality, not the case.

My feeling is it would be better to have a "detect" option that did not auto-allocate labels. I guess there's an argument to be made that it would take as many clicks to assign the correct labels as to unassign the incorrect ones, but once incorrect ones are assigned it's a real pain to go through and unassign; and it feels wrong to create an error situation in the first place - better an "opt-out" system than an "opt-in" (or at least a choice between these behaviours).
Snowman1
http://www.flickr.com/photos/snowman-1/
--------------------------------------

Hert
Posts: 6874
Joined: 13 Sep 03 7:24

Re: Google AI and PSU design

Post by Hert » 21 Jun 21 21:09

You don't mention how you do AI. But in the LAP you can click "Detect Labels" and that will not assign known labels.
You can click the hamburger menu in AI and select two different Detection Methods: assigning detected known labels or assigning all detected labels.
DetectLabels.png
DetectLabels.png (194.83 KiB) Viewed 354 times
An in the Batcher you can also specify how detected labels should be treated:
DetectLabels2.png
DetectLabels2.png (238.76 KiB) Viewed 354 times
Use what works for you
This is a user-to-user forum. If you need product support then please send a message

snowman1
Posts: 329
Joined: 01 Jan 07 3:13
Location: UK

Re: Google AI and PSU design

Post by snowman1 » 21 Jun 21 22:56

Ah, and does clicking the big "detect labels" button in the LAP - the one that's just behind the drop-downs in your first screenshot - does that always correspond to "detect and assign known labels"?
Snowman1
http://www.flickr.com/photos/snowman-1/
--------------------------------------

Hert
Posts: 6874
Joined: 13 Sep 03 7:24

Re: Google AI and PSU design

Post by Hert » 21 Jun 21 23:49

It always uses “none”
This is a user-to-user forum. If you need product support then please send a message

snowman1
Posts: 329
Joined: 01 Jan 07 3:13
Location: UK

Re: Google AI and PSU design

Post by snowman1 » 22 Jun 21 0:43

That's strange - I've checked several times and it's always behaved as per "detect and assign known labels". Regardless, now I know I can use "detect labels" from the dropdown (or the equivalent you've pointed out on the batcher) that's what I need, thanks for pointing me in the right direction.
Snowman1
http://www.flickr.com/photos/snowman-1/
--------------------------------------

Hert
Posts: 6874
Joined: 13 Sep 03 7:24

Re: Google AI and PSU design

Post by Hert » 22 Jun 21 8:24

Hert wrote:
21 Jun 21 23:49
It always uses “none”
I must correct myself here. Clicking the button indeed uses "Assign Known Labels"
This is a user-to-user forum. If you need product support then please send a message

snowman1
Posts: 329
Joined: 01 Jan 07 3:13
Location: UK

Re: Google AI and PSU design

Post by snowman1 » 22 Jun 21 11:35

No problem and thanks again for the help.
Snowman1
http://www.flickr.com/photos/snowman-1/
--------------------------------------

Post Reply