Analyzing Single Image 3D Reconstruction Methods: MarrNet and 3D Modeling using Reinforcement Learning

A single image 3D reconstruction method generates a 3D model given a 2D image. This research looks at two single image 3D reconstruction frameworks that have been proposed in literature: MarrNet and Modeling 3D Shapes by Reinforcement Learning (3DRL). Both methods use an intermediation approach where they break the problem into sequential steps, generating intermediate representations of the model.To further understand the importance and advantages of these steps, I design a set of experiments that compare both approaches from an analytical point of view. Additionally, I propose a hybrid framework inspired by both methods.