How to know if my detection model is overfitting?

Hi everyone,

I’m a bit new to the world of object detection and I’m wondering if there is a way to see if my model is over- or underfitting during training? Similar to how you would do it for a classification model.

I’m following the tutorial on finetuning a detection model: 08. Finetune a pretrained detection model — gluoncv 0.11.0 documentation

But I’m slightly confused on how the training and validation splits are used during the training of the model.
Both the cross entropy loss and the smooth l1 loss is printed during training, but is that for the training data or the validation data?

Is there a conventional/standard way to visualize overfitting for detection models?

Thanks!

Is there a conventional/standard way to visualize overfitting for detection models?

Usually you print 4 values at the end of your epochs:

  • train accuracy
  • validation accuracy
  • train loss
  • validation loss

overfitting is characterized by a divergence between the train and the validation accuracy. The validation accuracy stagnate while the train accuracy increases. This decoupling is a sign of overfitting.
Take a look to the following picture.
Taken from here.

From epoch > 6 this model starts to overfit.