Google AI - spectacularly bad!

Post Reply
snowman1
Posts: 273
Joined: 01 Jan 07 3:13
Location: UK

Google AI - spectacularly bad!

Post by snowman1 » 31 Oct 20 21:34

I haven't processed many images since setting up PSU for the AI options but just got round to importing the first batch of many on my memory cards. I'd set PSU up for AI scans on import. And wow is Google AI poor!

This particular batch of 225 pics were from the Yorkshire Dales so I was a bit surprised when it labelled (and geotagged!) one of them as being in Vietnam (Ngoc Linh Nature Reserve), one of them as on the Russia-Lithuana border (Nationaal Park Koerse Schoorwal / Curonian Spit) and a third as Buena Vista! (though all look great suggestions for places to visit!)

Just about everything got labelled as an animal (type unspecified), whether it was or it wasn't (mostly it wasn't: dry stone walls and boulders seemed especially likely to be labelled as animals). Landscapes with trees got labelled as plants which whilst technically true wasn't very helpful.

It did score some successes - some sheep got identified correctly, and in one picture it correctly identified that my other half was wearing shorts!

But overall the accuracy was very low, and where the label/area was accurate it wasn't very useful, and cleaning up the errors will take far more time than it would have done to add the few that (a) it got right and (b) were useful. So it's getting turned off.

I'm glad I tried this on a small batch at first.

Is this what everyone else has found or was I just unlucky on this batch? Does anyone have any recommendations for improving accuracy, or for when it should and shouldn't be used? I'd be very interested to hear other people findings, and whether different types of subject have got better results.

I must add that this is in no way whatsoever a criticism of PSU - it is an excellent move forward for PSU, definitely the right direction, and once the AI improves it will be extremely useful - it's just that Google's AI seems behind the point other services, e.g. Flickr, have got to. I'm sure it will improve but right now it doesn't seem to be where it needs to be to be useful.
Snowman1
http://www.flickr.com/photos/snowman-1/
--------------------------------------

User avatar
G8DHE
Posts: 246
Joined: 21 Aug 17 13:58

Re: Google AI - spectacularly bad!

Post by G8DHE » 01 Nov 20 18:35

I must admit I didn't have a lot of success, it was usually vaguely correct but tried often to be to accurate and promptly failed, unable to differentiate between the many based water sports at all accurately, if it had perhaps adopted a hierarchal approach it might have been useful e.g. Water Sports, paddle / wind / body it would have been useful but when it only gives a single identification i.e. Surfing (actually a Prone paddler), mixing up all the standing activities like wind surf, paddle board, kite surfing it becomes less than useful for me.
Geoff Mather (G8DHE)

Mke
Posts: 520
Joined: 15 Jun 14 15:39

Re: Google AI - spectacularly bad!

Post by Mke » 01 Nov 20 18:44

Well from the number of annoying CAPTCHAs I have to complete it must be very good at identifying fire hydrants, chimneys, traffic lights, cars, buses, bicycles and crosswalks.

Guess they've not moved on to boulders and dry stone walls yet :)

snowman1
Posts: 273
Joined: 01 Jan 07 3:13
Location: UK

Re: Google AI - spectacularly bad!

Post by snowman1 » 02 Nov 20 12:36

Mke wrote:
01 Nov 20 18:44
Well from the number of annoying CAPTCHAs I have to complete it must be very good at identifying fire hydrants, chimneys, traffic lights, cars, buses, bicycles and crosswalks.

Guess they've not moved on to boulders and dry stone walls yet :)
😂😂😂!!! They obviously have a long term master plan!
Snowman1
http://www.flickr.com/photos/snowman-1/
--------------------------------------

fbungarz
Posts: 1594
Joined: 08 Dec 06 5:03
Location: Arizona, USA

Re: Google AI - spectacularly bad!

Post by fbungarz » 03 Nov 20 1:19

As far as I have tried it so far it is completely useless...

Hert
Posts: 6426
Joined: 13 Sep 03 7:24

Re: Google AI - spectacularly bad!

Post by Hert » 03 Nov 20 10:02

I find it very useful. Maybe it's all about expectations. I wouldn't use it on import. But I like using it in the Label Assignment Panel. PSU/Vision then gives suggestions and I can select the relevant hits.
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

User avatar
Tom Husband
Posts: 85
Joined: 10 Dec 05 17:07
Location: Somerset, England

Re: Google AI - spectacularly bad!

Post by Tom Husband » 03 Nov 20 10:14

I've never tried using it on import but I find it pretty handy when assigning key words.

fbungarz
Posts: 1594
Joined: 08 Dec 06 5:03
Location: Arizona, USA

Re: Google AI - spectacularly bad!

Post by fbungarz » 03 Nov 20 18:46

Probably depends of what you are taking pictures off. I mostly do macro-photography of lichens on rock. I have played around with it and I get things like "rock" for a lichen growing on soil and "plant" for a lichen (which isn't a plant). I guess I never really expected this to work for my kind of a niche, so I am not really surprised that this won't work.

On the other hand: if you take pictures of the Eiffel Tower or Buckingham Palace I would guess it would be accurate. But those famous sites one does know anyway, right?
So, that perhaps leaves Google Vision tagging for general categorizing? Routinely automatically tagging cars as "car", fire hydrants as "fire hydrant", people as "people" might be quite useful. The question remains, what is more efficient: having to double-check if Google Vision recognized tagged images correctly or doing the tagging yourself. I guess I do not take enough general images to be the judge of that.

snowman1
Posts: 273
Joined: 01 Jan 07 3:13
Location: UK

Re: Google AI - spectacularly bad!

Post by snowman1 » 16 Nov 20 22:23

It does have the occasional unexpected success: one image of some stepping stones was tagged as river, body of water, stream, bank, and - I had to get my dictionary out to see what this was but it was entirely accurate - riparian zone. And a shot of a sea urchin was correctly labelled under marine invertebrate, sea urchin and echinoderm. But these seem to outweighed by entries that are either incorrect or correct but not in any way useful.

I think Hert's advice is sound, not use blanket-fashion on imports but selectively, where it will be most effective, on catalogued images.

Which brings me on to a question/request: on the right-click menu there is a script to remove areas for an image. But is there also a script that removes the linked labels (but not other labels) as well - is that possible?
Snowman1
http://www.flickr.com/photos/snowman-1/
--------------------------------------

Post Reply