Hello there,
I am trying to get started with Image classification which ultimately ist supposed to run on an esp32 cam board.
To get started i thought i could use a dataset from kaggle with a lot of images to be sure to have enough training data for the impulse creation.
But if i run live inference from my phone the correct label is picked up only very few times, so i thought i might ask here if there is a obvious mistake in this Impulse design or is it just not gooing to work well to use a external dataset and i should use the device/environment in which inference will run to collect training data (i was somehow hoping for a better result using the external dataset)
Two specific questions (but please point out any other points which you think are wrong)
-
in the impulse design Page, should i enter 96x96 if i use a 96x96 model, or is this the resolution of the Input/Training data (which ist 128x128)
-
how should i choose between the learning Block “transfer learning” vs. “Classification”?
This ist the project:
Thanks a lot for any remarks.