Computer Vision Research Project
Paper published in 2021 Neurips Workshop for Deep Generative Models and Downstream Applications
The goal is to train a model to reconstruct a face given a certain portion of a face. This portion is quantified as a patch, and the goal is to use as few patches to reconstruction the original image. A greedy approach is used to determine which patch locations are optimal for each image and the results are show down below.
The model consists of Unet architecture trained on a Perceptual Loss function based on VGG-16.
