Pretty cool, my little pony object detection first try

Well this was kind of fun

could not upload picture due to first post…

I ran this under object detection but i think it might be better to run it under classification, because well the only thing that changes is the little pony color, the shape is almost the same.

I got a precision score of 33 % with around 40 pics each for train and 25 as test data.

Maybe classification will work best ?


Your accuracy in float32 is way better (around 76%).
You can see this under Object Detection tab in the menu:
Screenshot 2021-08-03 at 08.48.15

And for object detection, default model using the edge-impulse-linux-runner is float32.

Feel free to also try with the image classification, models are lighter and you will be able to run it on smaller targets.



Hi @buton, as @louis said the quantization error can be big on object detection models (not such an issue on Linux devices at the moment), but we’ve had issues internally as well with objects that are the same shape but different colors - not entirely sure what the underlying issue yet. Two ways around it:

  • Image classification if you don’t need the spatial information.
  • Train a single object detection model (“my little pony”) and then just look at the most prevalent color in the box to distinguish between them.

The second approach is what @jenny did for a customer case on detecting things on a conveyor belt, works really well.


thanks @janjongboom and @louis

will keep experimenting for sure. but so far so easy to use, great job on the platform.