• State Resolved
  • Locked Locked
  • Replies 1 reply
  • Subscribers 98 subscribers
  • Views 142 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

Original question:

SK-AM62A-LP: Training failed during Image train with Edge AI studio Model composer

SK-AM62A-LP: Training accuracy is better but the false positive rate is very high

terry zhang
Prodigy 30 points
Part Number: SK-AM62A-LP

Tool/software:

Picture 1

We have completed the training and get the accuracy is 1（see Picture 1）, totally 3 sample: 1-bad,2-good 3-reverse, but the false positive rate is very high, especially for good sample,

Picture2

Rea box is table bad sample, blue box is table reverse sample and the other sample is good sample.

Could you give us any suggestion  to help us improve the recognition rate?

  • +1
    TI__Genius 15396 points

    Hi Terry,

    Since your model is at 100% on the training set but doing poor on real data / for testing, I have a feeling this is overfitting. It has learned the patterns in the training data so well that it is struggling to work on new data that doesn't look exactly the same.

    How many training images do you have? Overfitting is more likely to happen on smaller datasets and/or too many training epochs.

    Without knowing about your training dataset, I suggest capturing more images, especially ones with the 'good' class. 

    • It is smart to have various lighting conditions (intensity and placement of light source --> will impact shadows and reflections).
    • The location and orientation of objects is also good to vary, but YOLO-based models are generally robust to this