NeurVPS: Neural Vanishing Point Scanning via Conic Convolution

  • 2019-10-14 17:58:23
  • Yichao Zhou, Haozhi Qi, Jingwei Huang, Yi Ma
  • 9

Abstract

We present a simple yet effective end-to-end trainable deep network withgeometry-inspired convolutional operators for detecting vanishing points inimages. Traditional convolutional neural networks rely on aggregating edgefeatures and do not have mechanisms to directly exploit the geometricproperties of vanishing points as the intersections of parallel lines. In thiswork, we identify a canonical conic space in which the neural network caneffectively compute the global geometric information of vanishing pointslocally, and we propose a novel operator named conic convolution that can beimplemented as regular convolutions in this space. This new operator explicitlyenforces feature extractions and aggregations along the structural lines andyet has the same number of parameters as the regular 2D convolution. Ourextensive experiments on both synthetic and real-world datasets show that theproposed operator significantly improves the performance of vanishing pointdetection over traditional methods. The code and dataset have been madepublicly available at https://github.com/zhou13/neurvps.

 

Quick Read (beta)

loading the full paper ...