Edge impulse gives the ability to train a model using two labeling methods, 1. one label per item and 2. bounding boxes. Can I train a model using the first labeling method and then train that same model later on using the second method?
I have trained a model for digit recognition(0-9) using only one label per data that has a 96% accuracy but when the model is deployed it has a hard time reading the number and I am assuming is because it doesn’t know where the number is in the frame.
I agree with your assumption. If the training data uses digits that take up the whole frame, then the deployed model will have a hard time prediction a digit that only takes up 1/2 of a frame. If this is the case then you’ll need to add more training data. For example, large digits, medium, digits and small digits. Also, make sure to add rotated digits as well.
You can change your Model from a Classification Model to a FOMO Model.
If you want to take a FOMO Model and make it a Classification Model, then you need to download all the data, delete the bounding boxes file (in the downloaded data), delete all the data in the EI Studio, upload the files to EI Studio.
There is a lot of things to consider when deploying to the hardware, it would be ideal if you can tell us more on what device you have deployed the model to.
the FOMO object detection model is what we call “fully convolutional” which means you can train at a different resolution to what you run inference at. specifically for your example you could train at a small resolution in a way equivalent to one label per item and then run the model at a larger resolution giving effectively bounding boxes… would take a bit of wrangling though since it’s not something directly supported.
do you expect the digits to occupy the same area of pixels between your one label per data case you’ve already trained and the bounding box case?