Google Cloud Vision API test

Index, feed.

Now that Google provides a "Vision API" cloud service, to annotate pictures in several ways. The "label" type seemed especially promising, and I thought I would try it on some of my photographs to see if it did a better job than lazy-me captioning by hand.

The above image, depicting Gabriel Nadeau-Dubois and Jeanne Reynolds giving a press conference under the rain on 2012-06-02, was labeled as follows:

{
  "labelAnnotations": [
    {
      "mid": "/m/07yv9",
      "description": "vehicle",
      "score": 0.71177125
    },
    {
      "mid": "/m/09g5pq",
      "description": "people",
      "score": 0.70403248
    },
    {
      "mid": "/m/01hr1w",
      "description": "guided missile destroyer",
      "score": 0.54956192
    }
  ]
}

It's a good thing that Google's artificial intelligence was able to identify those dangerous terrorists' guided missile destroyer, because I certainly couldn't.

In all seriousness, the auto-labeling feature generally outputs relevant labels, although it sometimes misses things that are obvious to a human observer. In this specific case, note that the AI was only about 50% confident that it had identified a guided missile destroyer; perhaps it's more annoying that it identified a non-existent vehicle with 70% confidence.

www.kurokatta.org

David Haguenauer <www@kurokatta.org>
$Id: google-vision-wmd.malblog,v 1.6 2016/04/11 21:21:22 homer Exp $