Group Actions
Engagement
  • Author Author: ntewinkel
  • Date Created: Date Created
  • Views 609 views
  • Likes 8 likes
  • Comments 3 comments
Related
Recommended

Honey Bee Safe: Nicla Vision Again

ntewinkel
Honey Bee Safe: Nicla Vision Again

I thought I'd give machine vision and identification on the Nicla Vision another try, as last time things went terribly wrong, with unicorns everywhere!

This time I focused on just identifying rubber duckies, without confusing the model with the added unicorns. I also added a “background” class for when there is nothing to be seen. I’m not sure if it also needs a class for unknown things.

I took 100 pictures of the 3 rubber duckies I have, from various angles and with different lighting, distances, and shadows thrown in.
I also took 30 pictures of the background with various lighting angles, reflections, and shadows, with some pictures showing the edge of the desk as well.

ducks photo shoot

Then I erased all of the data in my Edge Impulse project and exported the new dataset from OpenMV again to restart the training process.
The model said it was getting 86.4% accuracy, and it had a bunch of green dots for correctly identified ducks.. but none for correctly identified backgrounds Thinking

edge impulse results

Then I noticed at the top it said something about “Target: Arduino Portenta” hmm… I should switch that to the Nicla Vision board. They appear to have similar chip and speed though.
Then, as before, in the Deployment section, I built the firmware again for OpenMV and downloaded it to my computer.

edge impulse target board

As per my previous blog post, I then uploaded the firmware and ran the example script to test out the new Duck Detector.

This time things went much more smoothly!

Ducks were spotted successfully!!

duck spotted seriala duckanother duckyet another duck

And no ducks when there was no duck!

no duck!

So by now I was feeling pretty good about myself and how things were working well this time. until...

iPhone is a duck

The iPhone is a duck!
Well dang. ok there are pictures of ducks on my iPhone, but that doesn't count!

A little more testing confirmed my sad suspicions - that anything... ANYTHING... in the field of view is identified as a duck. Anguished

Is there some way to have an "unknown" class where everything else is dumped that doesn't match?

If anyone has ideas on how to overcome the "anything is a duck" issues, I'd love to hear your suggestions!

  • in reply to shabaz

    I think I figured it out last night - the model I was using, which was what they show in the Arduino example, is to classify objects, not really to identify or count them.

    By that I mean, similar to what you said, it is only expecting a known set of objects to show up - like on an assembly line.

    By using a "bounding boxes" model I was able to detect a duck, a few ducks at once, as well as stuff that's not a duck.

  • in reply to shabaz

    (Can't edit). By "the training data" I also mean the test data too.

  • If it walks like a duck.. : ) 

    I've not used Edge Impulse much (I gave it a quick shot when they went live), but I think there is a  problem in the training if it only found ducks in that chart in that app. The training data (if all you're interested in is ducks) should be "ducks" and "not ducks", i.e. the latter would include the background and everything else. In other words, the training data would include things like mobile phones and other random objects in there, all with the "not ducks" label. I'm guessing (I'm not an expert) that the particular model has incorrectly learned that anything that is not a background is a duck, which to be fair, is what it was trained with so it's working as expected.