We talked before about the minimum score for highlighting. Let's see exactly what that means by taking a look at what happens when we use different minimum scores for object highlighting. Let's start out with a value of 0.5 and see what objects are detected within our photograph:

As you can see, we have two kites selected with a fairly good accuracy score attached to each. The boxes are drawn in green to indicate high confidence targets. Not bad. But there are still a lot of objects out there that I think we should be picking up. There are a few more kites and several people that should be easy to detect. Why haven't we done so?

What if we lowered our minimum threshold to 0.3 instead of 0.5? Let's see the result:

Well, as you can see, we do pick up other kites, albeit with a lower confidence score due to their distance in the photograph, but we have also, more importantly, now started to recognize people. Any box drawn in red is a low confidence target, green is high, and yellow is medium.

Now, what if we went one step further and lowered our minimum threshold all the way down to 0.1? If our pattern follows, we should be able to identify more images, albeit with lower confidence scores, of course.

If you look at the following version of the photograph, you can see that we do, in fact, have many more objects selected. Unfortunately, the accuracy has diminished considerably, as we suspected. Kites were confused with people, and in one case a tree was confused with a person as well. But the positive note is that our recognition changes as we adjust our threshold. Could this be done adaptively in a more advanced application? Absolutely, and it's those kinds of thoughts I want to nurture so that you can embellish the code and make truly earth-shaking applications:

OK, there's one final example that I think you will like. In this example, I have dropped the minimum threshold down to 0.01. If our hunch is right, the screen should light up with low confidence targets now. Let's see whether we're right:

It looks like our hunch was right. I know the screen labeling is cluttered, but the point is that we have increased our object detection, albeit for a lower confidence threshold.

You should now take some time and consider all of the exciting applications for such technology. From face and object detection to autonomous vehicles, Tensors are used everywhere today and it's something you should get familiar with.