Supervisors: Dr. Ajad Chhatkuli, Thomas Probst, Prof. Luc Van Gool
Vanishing points are useful for several tasks in 3D vision such as computing camera intrinsics and extrinsics as well as scene understanding. However detecting them is difficult and usually requires several steps such as detecting lines, removing outliers and classifying them according to the orthogonal direction in a classical pipeline. Although this could be simplified and speeded-up using a machine learning pipeline, it is not clear how geometric constraints can be included in the learning mechanism. End-to-end learning of vanishing point detection therefore remains unsolved. A key component here is the line detection from edges. Previous works accomplished the task using non-differentiable algorithms such as the Line Segment Detector (LSD), though the latter task of computing the vanishing point was solved using a deep neural network. In this work, we tackle the end-to-end vanishing point detection through the supervised and unsupervised line detection methods applied on CNN based edge detector outputs. The end goal is to compute the vanishing point from the detected Manhattan lines. We show some positive and negative results obtained thus far in the project.