IssueTracker

The new label detection model released in: https://cloud.google.com/vision/docs/release-notes#August_23_2024 seems to cause quite a radical difference in the labels that are returned.

I have attached a copy of an image to reproduce the issue. Sensitive data has been blurred.

Problem you have encountered:

• Running the attached image with builtin/legacy will return the labels: Hood,Automotive lighting,Gesture,Automotive mirror,Bumper,Automotive exterior,Gadget,Vehicle door,Gas,Netbook
• However against the current stable model we get no returned labels at all (Even running the image through without any blurring).

What you expected to happen:

• At least some annotations to be returned.

Steps to reproduce:

• Run image through where the model property of the features object is builtin/legacy. Observe labels are returned.
• Run image through where the model property of the features object is omitted so it uses the current stable model. Observe no labels are returned.

Perhaps there could be someway to continue using older models through the API? as it looks like you won't be able to use the legacy model anymore after a further 90 days.