3 skills found
ParthaEth / GIFGIF is a photorealistic generative face model with explicit 3D geometric and photometric control.
pandaypr / Interpretation Of Activation Maps In Generative ModelsRecent work in the field of Explainable AI and Computer Vision on CNN based architecture has improved the interpretability of Deep Learning models and helped in visualizing the model pre- dictions. Methods like CAM, Grad-CAM and Guided Grad-CAM have proved the practicality of localized visual attention in the classification and categorization applications. However, not much research has been done on generative models. In our work, we implement Grad-CAM technique on VAE and CVAE models trained on CelebA-HQ dataset and calculate neural attention map. The aim of the project is to build generative models capable of generating controllable human faces and build semantic segmentation of human face, and then investigate methodologies to improve the explainability by applying Explainable AI tech- niques like Grad-CAM, and analysing the effect of altering the model architecture, loss functions, latent space size. Furthermore, we investigate the latent space information of models by modifying the latent node variables.
adp178 / Interpreting The Latent Space Of GANs For Semantic Face Editing This project aims to build a controllable Generative Adversarial Network (GAN) that facilitates the control of facial attributes through manipulation of latent space. We have used pre-trained StyleGAN2-ADA to generate photorealistic faces and SVM hyperplanes to interpret it’s latent space. This allowed us to understand the encoding of facial semantics and their disentanglement in the latent space learnt by GAN.
