3D Reconstruction of Multi-View Flat Sketches

Abstract

We propose and implement a system for 3D reconstruction of flat sketches. Our system takes two sketches of two views of the shape as an input and outputs 12 corresponding multi view images with normal and depths maps. Architecture of the system is composed of two components which are Generator and Discriminator. Generator is responsible for generating the 12 images from the sketches while discriminator’s objective is to fake images generated from generator and real images which are the targets. Generator succeed when it manages to fool discriminator into classifying generated images as real ones. Generating images is accomplished through two steps: (i) encoder which convert input sketches into a feature map representation, and (ii) decoder which map this feature representation into images. Discriminator is a linear classification convolutional network which classifies images as real or fake. Our system produced nearly perfect multi-view images from detailed and undetailed synthetic sketches, and from hand-drawn sketches.