Trying TensorFlow’s image recognition script
For each image, TensorFlow returns a list of 5 keywords along with a score indicating a level of confidence. With this gorilla picture, the result is very clear.
Some mistakes are quite understandable, like when it hesitates between an ox and an ibex to describe this yak.
It does fail sometimes, like for this violet-backed starling which cannot be identified. In these cases, TensorFlow only returns labels with a low confidence score.
Scores tend to get lower when the picture contains multiple subjects, like in this picture from Noah’s Ark Animal Sanctuary.
I tried to isolate the subjects to understand if TensorFlow was struggling with one of them. Submitted individually, it identifies the bear without any doubt. I suspect that it’s not listed with the original image because of the 5 keywords limit.
TensorFlow is sometimes surprisingly good at identifying something from a small part of it. This picture is one of the best examples that I could find:
If we crop the image to exclude most of the animal, TensorFlow remains highly confident about the fact that there is an armadillo in the picture.