zhangyhcumt / image_segmentation Goto Github PK

This is a image semantic segmentation demo using Keras.
To simplify the code, I choose the horse dataset, as the two classes are quite balanced (background and horse).

Horse dataset is downloaded from http://www.msri.org/people/members/eranb/

1. FCN, https://arxiv.org/abs/1411.4038, translated from the original caffe code https://github.com/shelhamer/fcn.berkeleyvision.org
2. Unet, https://arxiv.org/abs/1505.04597
3. DeepLab V3+ (onging), https://arxiv.org/abs/1802.02611

Inside the custom_loss.py, there are some losses not only for segmentation task, also for binary classification or category classification. Some famous loss implemented, such as focal loss.

The custom_loss_eagermode.py is only for loss function testing purpose, testing on eager mode is more efficient.

However the class imbalance is always a big problem in daily segmentation tasks.
Tried on Pascal dataset, but the result is bad, still exploring.

1. Weight cross entropy, what's the reasonable loss weights for classes ? (the inverse class frequency? ongoing)
2. Dice loss / GDL (onging)
3. Tversky loss (onging)