Visual Boundary Prediction: A Deep Neural Prediction Network and Quality Dissection

Abstract

This paper investigates visual boundary de-tection, i.e. prediction of the presence of a boundary at a given image location. We de-velop a novel neurally-inspired deep architec-ture for the task. Notable aspects of our work are (i) the use of “covariance features” [Ranzato and Hinton, 2010] which depend on the squared response of a filter to the in-put image, and (ii) the integration of im-age information from multiple scales and se-mantic levels via multiple streams of inter-linked, layered, and non-linear “deep ” pro-cessing. Our results on the Berkeley Segmen-tation Data Set 500 (BSDS500) show com-parable or better performance to the top-performing methods [Arbelaez et al., 2011, Ren and Bo, 2012, Lim et al., 2013, Dollár and Zitnick, 2013] with effective inference times. We also propose novel quantitative assessment techniques for improved method understanding and comparison. We care-fully dissect the performance of our architec-ture, feature-types used and training meth-ods, providing clear signals for model under-standing and development. 1