-
|
I understand that the background category is to 0 by default. Because the number of negative images (image with no annotations) is much larger than image with annotations, my pipeline required learning from negative images. Best regards |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments 1 reply
-
|
@hieutrluu You shouldn't need to do anything specific for that case, just leave those images with no bbox annotations. The losses should take care of penalizing any predictions for negative images. And the loss config is already designed (retinanet style) to take into account the fact that easy negatives should be downweighted since that's the majority of image space regardless of whether or not you have lots of negative images. The only thing you may need to do if you have an abnormally large imbalance, is to tweak tweak some of the loss weightings/params |
Beta Was this translation helpful? Give feedback.
-
|
thank you very much for your help and amazing repo |
Beta Was this translation helpful? Give feedback.
-
|
@rwightman I am not sure I understood your response. What should be the input to the model during training when a batch has both positive and negative images? |
Beta Was this translation helpful? Give feedback.
@hieutrluu You shouldn't need to do anything specific for that case, just leave those images with no bbox annotations. The losses should take care of penalizing any predictions for negative images. And the loss config is already designed (retinanet style) to take into account the fact that easy negatives should be downweighted since that's the majority of image space regardless of whether or not you have lots of negative images.
The only thing you may need to do if you have an abnormally large imbalance, is to tweak tweak some of the loss weightings/params