Inferring 3D Scene Structure from a Single Still Image

In this project, we revisit the problem of constructing 3D structures from single still images. We build upon previous work done by Saxena, Sun and Ng [1] by improving on the inference techniques used in their algorithm, with the goal of producing 3D models that are more quantitatively accurate, as well as more visually pleasing. One area of improvement in the existing algorithm is the penalty function used during MAP inference of plane parameters. When inferring 3D models from single still images, the penalty function is used to enforce constraints such as connectedness, co-planarity, and co-linearity. Properties of the penalty function consequently determine whether the transition between two planes in the 3D model is smooth or sharp. The current penalty function, the L1 norm of the error, does not prefer either a smooth transition or a sharp transition. As a result, the resulting 3D models often have walls sloping away from the ground, rather than standing straight up. Our goal was to find a suitable penalty function that prefers a sharp transition over a smooth transition.