Semantic Segmentation using Adversarial Training

Paper Ref :- Semantic Segmentation using Adversarial Networks

Pytorch implementation

Stanford Background Segmentation dataset is used. There are 8 classes. The label maps are one-hot-encodified and compressed and stored in HD5F format.

Loss functions:

Discriminator's Loss

Real_Loss = BinaryCrossEntropyLoss(discriminator(Real_Label_Map), Ones)
Fake_Loss = BinaryCrossEntropyLoss(discriminator(Fake_Label_Map), Zeroes) ## Fake label map => generated by the segmentor

Segmentor's Loss

SLoss = CrossEntropy(Generated, Actual) + Lambda*BinaryCrossEntropyLoss(discriminator(Fake_Label_Map), Ones)

Prolonged access to GPU was not available. This was trained for 12 epochs with only 1 image per mini-batch ( lack of high memory GPU ).

The results are:-

Losses:

Discriminator's fake loss

Discriminator's real loss

Segmentor's Loss

Class 1 :

Class 2 :

Class 3 :

Class 4 :

Class 5 :

Class 6 :

Class 7 :

Class 8 :