Machine Vision Solved (Almost)

September 11, 2014

I read “The Revolutionary Technique That Quietly Changed Machine Vision Forever.” The main idea is that having software figure out what an image “is” has become a slam dunk. Well, most of the time.

The write up from the tech cheerleaders at Technology Review says, “Machines are now almost as good as human at object recognition.”

A couple of niggling points. There is that phrase “almost as good”. Then there is the phrase “object recognition.”

Read the write up and then answer these questions:

  • Is the method ready to analyze imagery fed by a drone to a warfighter during a live fire engagement?
  • Is the system able to classify a weapon in a manner meaningful to field commander?
  • Can the system discern a cancerous tissue from a non cancerous tissue with an image output from a medical imaging system?
  • Does the method recognize objects in a image like the one shown below?

P1010077

Image by Stephen E Arnold, 2013

If you pass this query to Google’s image recognition system, you get street scenes, not a person watching activities through an area cordoned off by government workers.

image

Google thinks the surveillance image is just like the scenes shown above. Note Google does not include observers or the all important police tape.

The write up states:

In other words, it is not going to be long before machines significantly outperform humans in image recognition tasks. The best machine vision algorithms still struggle with objects that are small or thin such as a small ant on a stem of a flower or a person holding a quill in their hand. They also have trouble with images that have been distorted with filters, an increasingly common phenomenon with modern digital cameras.

This stuff works in science fiction stories, however. Lab progress is not real world application progress.

Stephen E Arnold, September 11, 2014

Comments

Comments are closed.

  • Archives

  • Recent Posts

  • Meta