NIA image recognition software has learned thousands of species but still classifies humans as Highland Cattle

Highland Cattle - Bos taurus var. highland
with 100.0% probability.

Highland Cattle - Bos taurus var. highland
with 100.0% probability.
©https://thispersondoesnotexist.com

Humans have a lot in common with Highland Cattle! Let me explain...

NIA (Nature Identification API ) is the image recognition software that helps identify animal and plant species at observation.org. It all started in 2019 as ObsIdentify and was introduced on the Dutch website waarneming.nl. It is also available as an app. The purpose is automatic identification of animals and plants. The identification of a species is always presented as a prediction, it is never the final answer. A human moderator always makes the final decision. It is especially useful for inexperienced users (and fun for the experienced users). Since its appearance NIA was trained with a growing database of thousands of validated pictures. One would expect that its success rate would steadily increase. which is probably the case. But, I don't yet have definitive data about NIA's performance. However, I have noticed that its predictions have changed in the past few years.

However, unexpectedly, it is still easy to fool the software with pictures of human faces. For example the above computer generated image of a person is classified as Highland Cattle with 100.0% probability. I find that fascinating. How is it possible that this successful software makes such weird errors? Naturally, it sometimes confuses ponies or horses with highland cattle, but human faces? They are in a completely different category. What feature in the human face could be the cause of these mis-identifications? Maybe hair? Compare these two images:

Highland cattle 19.9%
©https://thispersondoesnotexist.com

  

Highland Cattle 98.0% (modified picture)
©https://thispersondoesnotexist.com

Clearly, duplicating part of the girl's hair enhances the confidence of the 'Highland cattle' prediction from 19.9% to 98.0%. It doesn't matter that it is a distorted face. The specific pattern and color of the hair is enough. 

Apparently, other people find this fascinating too. Some people do not only test their selfies, but also unnecessarily save them to the database. The system does not reject human faces automatically, simply because it thinks they are animals. I found the following pictures among the Photos of the 'Highland Cattle' of observation.org:

Highland Cattle 24,2%

Highland Cattle 79.1%

Highland Cattle 33.0%

Highland Cattle 27,6%

These four pictures are identified as Highland Cattle albeit with low probability, except for the dog. Obviously, they have not yet been seen by moderators. Otherwise they would have been rejected, I guess.

We humans have a lot in common with Highland Cattle! The pictures are not hard to find. The website thispersondoesnotexist.com produces an infinite number of human faces. Many of them are identified as animals. Some of them with high certainty. Endless pleasure to find them.

I found a very carefully executed time-lapse of a young women with one picture every week during 2 years:

Hair Growth Time-lapse - 2 Years (youtube)

This is really a great idea, but the most surprising thing is that every picture is classified as 'Cattle' or 'Highland Cattle' by NIA (observation.org):

Cattle - Bos taurus 81.9%

NIA image recognition predicts Highland Cattle
- Bos taurus var. highland with 99.4% probability.

I tested 10 pictures of the time-lapse: all are 'Cattle' or 'Highland Cattle' with probabilities ranging from 32% up to 99.4%. Again: we humans have a lot in common with Highland Cattle! For example: hair.

The NIA algorithm makes mistakes a child would not make. So, is this a stupid mistake? No, not really. It just views the world differently. And, it hasn't been trained to recognize humans! It is very good at what it has been trained to do: recognizing wild animal and plant species. NIA would only be in serious trouble when it identifies a wrong animal or plant species with 100% certainty.

Most humans are very good at recognizing human faces. The average person can recognize 5000 human faces and superrecognizers up to 10,000 humans. But most humans are very bad at recognizing wild animal species. Exceptions are top observers such as Remco Hofland with 8694 species. At the Dutch website waarneming.nl top observers Ruud en Betty van Middelkoop have a list of 9522 species. So, I guess, NIA / ObsIdentify most 'know' at least 10.000 species...

 

 

For more pictures see this page!

 Update 12 Dec: small edit in text.

 

Other blogs about NIA

• click on the blog label ObsIdentify
• Frequently asked questions about ObsIdentify  (Observation.org)
• I am not aware of any statistics of the number of species ObsIdentify is able to recognize.
• Amazon has 'celebrity recognition feature' to recognize celebrities in images (faces) and claims it 'recognizes tens of thousands of well-known personalities in images and videos'.