You can take a look at my
previous response posted for a similar question. If you are seeing differences for LABEL_DETECTION requests using the same API version, I ask that you report your findings in a
Public Issue Tracker with the image you used, the API requests sent, and the different response results so that the Vision engineering team can further investigate.