Happens also with cats and other animals
This is a new issue of the 'AI' model it does not happen on my A1 just on the A7C II but you have the A7C which I thought would not even do animals?
Yes, it has animal eye AF, just not all the fancy modes for different types of animal or the AI stuff.
I guess the ear instead of an eye is a retained feature on the advanced AI models as my A7C II detects it more easily than my A1
I don't Sony subject detection is very smart but their autofocus is good so once it locks on the ear you get a very sharp ear shot...
Honestly every time something like this happens I get less worries about AI’s ability to take my job!
I don’t shoot many statues or dogs, so haven’t seen this behaviour on my A7C.
Easy to blame the technology this is a deep learning model that has been fed with images and told which one was and which not then what it does is to look up some shapes on the catalog and come up with yes no
What you just described is actually how pre-AI object recognition worked. Literally pairing up images and then doing a numerical analysis on how close it is to said picture. The AI doesn't have a notion of a database of what an eye looks like - instead it's a pattern. This reason is why the AI model doesn't necessarily get bigger the more images you feed it.
as ears on cats and dogs are on the top of the head and eyes in the middle this means this algorithm is not astute and therefore it looks on the most prominent circular shapes which in animals can be inner ear
The AI model (at least in pre-A7RV days) has no notion of what the head is, or that an eye should belong where the head is for anything besides humans. It's why there's a lot more false positives with the older way of doing things - there's no semantical rules.
In a way I'd actual prefer that there be more discrete selection of what animals, and tied together with a generalised model that can automatically determine what the subject actually is first then selecting the best model. The A1II's way is paving the way for such a method.
if it checked that the eyes cannot be the highest point of the head it would not come as false positive but the routine is keen to lock early so as soon as it sees two dark holes it calls it success
Or that the AI model thinks that region is more likely that it's an eye than the actual eye itself.
Sorry but thats incorrect machine learning models are not very dissimilar to previous implementations a nd human face detection was already able to discern all of the parts of a face before however the processing would be done from scratch in camera with a decision tree this is heavy and therefore older models did just humans as computations become complex
The logic now is pre wired but essentially has just become faster as the camera looks up results
the two cameras a7c and a7c II make the same wrong call but the A7C Ii makes that call quicker and with a smaller subject accuracy wise I can see there is no difference between the camera generations
what the new ones do is to detect an object shape partially obscured however in terms of locking on the eyes the accuracy seems identical I have ran several scenarios to compare my A1 and my A7C II and indeed the A7C II gives positives at distance when you select head and body however if you choose eye only it has only in some cases a marginal benefit in how far away it recognises and mostly on static subjects
Dont want to go too off topic but the eats instead of eyes is an issue also on current range
I wasn't talking about the difference between the A7C and A7CII, I know its another inference model they're running, the main difference is the availability of more complex models powered by a dedicated chip. I was talking about your description of a lookup catalog.
Either way, it's a minor nitpick, don't wanna derail this thread so I won't go further.
Shame to hear that about the new AI models - I don't have an issue with ears, but with birds sometimes the dark spot behind a bird's actual eyes are misidentified.
The ear issue happens with the Animal setting. According to sony documentation
https://support.d-imaging.sony.co.jp/support/ilc/autofocus/ilce1/en/animaleyeaf.html
When shooting in dark places or animals with dark hair
the camera would have a challenge. My cat has grey fur which I would not sure if classifies as dark
the A7C II at distance detects the ear as eyes in broad daylight. The A1 detects nothing at all
When the cat is near the A1 and the A7C II detect the eyes correctly
I have not tried with a statue as the example here however my dark hamster that is dark and you could argue does not have a dog or cat face is detected by the camera but it goes on the ears with both
Birds have more complicated cases it looks like the camera only detects certain type of birds shape and if the beak is different it would struggle
In general the camera does not really recognise animals it really recognises cats and dogs. When you have an elephant or a bear standing on two feet it will struggle as per their example
The animal setting really is a pet setting, lion and tigers look like cats or it would not easily recognise them either
