to modify loss function for unbalanced dataset

matterport / Mask_RCNN

Mask R-CNN for object detection and instance segmentation on Keras and TensorFlow

Other

24.69k stars 11.71k forks source link

to modify loss function for unbalanced dataset #1189

Open nirajvermafcb opened 5 years ago

nirajvermafcb commented 5 years ago

I am able to train the model successfully but my problem is my dataset is quite unbalanced and has 30+ classes out of which 1 class is quite dominating and some classes are quite rare.I want to know whether unbalanced dataset will affect model accuracy and predictions.If yes except undersampling or upsampling is there any techniques I can do. Is there a way to change loss function which will penalise classes as per their distribution

eralmual commented 5 years ago

Were you able solve it? I'm facing the same issue

windson87 commented 5 years ago

me too

fuaiguo commented 5 years ago

Were you able solve this issue? @nirajvermafcb

ChienWong commented 4 years ago

use focal loss，Maybe improve it

charlie6echo commented 3 years ago

Here is my implementation of Multi-class Focal loss on Mask RCNN for Imbalance Dataset.