Google Cloud Vision OCR API returning incorrect values for bounding box/vertices

Question

I'm using the "TEXT_DETECTION" option from the Google Cloud Vision API to OCR some images.

The bounding box around individual characters is sometimes accurate and sometimes not, often within the same image.

Is this a normal side-effect of a probabilistic nature of the vision algorithm, a bug in the Vision API, or of course an issue with how I'm interpreting the response?

Here's the portion of the response specific to the letter "a" from which I'm extracting the bounding box.

stdClass Object
(
    [property] => stdClass Object
        (
            [detectedLanguages] => Array
                (
                    [0] => stdClass Object
                        (
                            [languageCode] => en
                        )

                )

        )

    [boundingBox] => stdClass Object
        (
            [vertices] => Array
                (
                    [0] => stdClass Object
                        (
                            [x] => 419
                            [y] => 304
                        )

                    [1] => stdClass Object
                        (
                            [x] => 479
                            [y] => 304
                        )

                    [2] => stdClass Object
                        (
                            [x] => 479
                            [y] => 397
                        )

                    [3] => stdClass Object
                        (
                            [x] => 419
                            [y] => 397
                        )

                )

        )

    [text] => a
)