In this project, we tackled the classical image Segmentation problem, from the perspective of image to image transaltion. We developed a generative neural architecture that performed semantic segmentation.
Further, we introduced the Gradient Loss (A novel implementation) to localize and preserve edges in the segmentation map. We observed that the introduction of this loss led to faster convergence.
In order to study and interpret the Network, we performed a PCA analysis followed by a GMM clustering of the encoder's logits. For selection of the optimal number of clusters, we introduced the Combined Information Criterion (CIC).
Combined Information Criterion:
Training Progression
