Vision training.

Robotics & Edge Computing Jetson & Embedded Systems Jetson Nano
1

Hi all

I am new to AI but really enjoying it so far…

I have been through the tutorials for 48 hours to a demo, i have completed cats and dogs and also plants.

I have tried to train my own with a bottle. I wanted to be able to detect good bottle vs bad bottle. IE. bottle stood up, bottle fallen over.

I have set it up, taken photos of about 200 good and 200 bad, ran program for 35 epochs but it the vision just sees a bottle and says good bottle.

It cannot dfetect the orientation.

What is the best method for trainig vision to detect object orientation. Can you explain how i should go about this for the best results? Number of pictures, number of epochs, type of pictures required.

Should i use something else??

Any tips would be most appreciated.

Thanks

Mark

2

You might want to “stretch” your 400 pictures, but giving them slight rotations, skews, flips, and so forth, and still keep the appropriate label. You’ll do better in training with more input data.

Also, you say you trained 35 epochs, but what loss did you get? If the loss doesn’t go down, then something is wrong with your dataset, your labels, your training hyperparapemeters, or your model. You should graph the progress of the model over time.

3

ok thanks, so do you think it is possible top detect object orientation by only visual training only?

Do i require some additional software?

Thanks

Mark