2015-471 – Constrained-CNN for 3D Scene Understanding

In the past years Convolutional Neural Networks (CNNs) have shown incredible promise in learning visual representations. In this work, we use CNN for the task of surface normal estimation. We propose to build upon decades of hard work in 3D Scene Understanding to design new architecure for surface normal estimation. Our architecure incorporates several constraints (manhattan-world, man-made) and several intermediate representations (edge labels, room layout). This architecure leads to state of the art performance on surface layout estimation task. Scott McEvoy smcevoy@andrew.cmu.edu 412-268-6053