I am setting up an application to use Google Cloud Vision API, and so far, I find the returned bounding boxes to only be the entire image. If for example the image is 100x100, then the bounding box is [(0,0), (100, 0), (100, 100), (0, 100)]. Even if there are multiple detected landmarks (and they are well separated in the image), they are all bounded by the entire image.

I even used the example image gs://cloud-samples-data/vision/landmark/st_basils.jpeg, and in the document it says the output is

{
  "responses": [
    {
      "landmarkAnnotations": [
        {
          "mid": "/m/014lft",
          "description": "Saint Basil's Cathedral",
          "score": 0.7840959,
          "boundingPoly": {
            "vertices": [
              {
                "x": 812,
                "y": 1058
              },
              {
                "x": 2389,
                "y": 1058
              },
              {
                "x": 2389,
                "y": 3052
              },
              {
                "x": 812,
                "y": 3052
              }
            ]
          },
          "locations": [
            {
              "latLng": {
                "latitude": 55.752912,
                "longitude": 37.622315883636475
              }
            }
          ]
        }
      ]
    }
  ]
}

But when I actually submit the request, I got

{
  "responses": [
    {
      "landmarkAnnotations": [
        {
          "mid": "/m/0hm_7",
          "description": "Red Square",
          "score": 0.7341708,
          "boundingPoly": {
            "vertices": [
              {},
              {
                "x": 2487
              },
              {
                "x": 2487,
                "y": 3213
              },
              {
                "y": 3213
              }
            ]
          },
          "locations": [
            {
              "latLng": {
                "latitude": 55.753930299999993,
                "longitude": 37.620794999999994
              }
            }
          ]
        }
      ]
    }
  ]
}

which is the entire image again.

I then watched this tutorial , and when he ran the code live, I can see some good result, but when I tried with the same image, I got nothing detected.

Has there been some significant change to the Cloud Vision API? It seems to do a lot worse than how it used to do. Is there any setting or anything I can do to improve the bounding boxes?