Deep Hough-Transform Line Priors

ChainLineNet: Deep-Learning-Based Segmentation and Parameterization of Chain Lines in Historical Prints

“…The holistically-attracted wireframe parser (HAWP) [11] was built on the L-CNN and introduced a novel line segment reparameterization by using a holistic attraction field map that assigns each pixel to its closest line segment. Lin et al [12] proposed in their deep Hough transform line priors method to combine line priors with deep learning by incorporating a trainable Hough transform block into a deep network and performing filtering in the Hough domain with local convolutions. For the application of line detection on the Wireframe datasets, they used the L-CNN [10] and the HAWP [11] as backbones and replaced the hourglass blocks with their Hough transform blocks.…”

“…During inference, the UNet was executed patchwise, and two postprocessing methods were applied to the reassembled segmentation output [6], which we refer to as PatchUNet-RANSAC and PatchUNet-Hough. Secondly, we trained the deep Hough transform line prior method [12] for our line detection task, which we abbreviate as PatchDeepHough. The method was originally developed for wireframe detection; thus, some modifications were necessary to make it applicable to our task.…”

Semi-Supervised Lane Detection With Deep Hough Transform

“…We encode an input image to a semantic feature representations F which is mapped to the Hough space, through a trainable Hough Transform module [9]. The Hough transform HT maps a feature map F of size [H ×W ] to an [N ρ ×N θ ] Hough histogram, where N ρ and N θ are the number of discrete offsets and angles.…”

“…To make effective use of additional unlabelled data, we propose a semi-supervised Hough Transform-based loss which exploits geometric prior knowledge of lanes in the Hough space [8,9]. Lanes are lines, thus we propose a semisupervised Hough Transform loss that parameterizes lines in Hough space, by mapping them to individual bins represented by an offset and an angle.…”

“…Lanes are lines, thus we propose a semisupervised Hough Transform loss that parameterizes lines in Hough space, by mapping them to individual bins represented by an offset and an angle. Inspired by the work in [9], we rely on a trainable Hough Transform and Inverse Hough Transform (HT -IHT ) module embedded into a neural network to learn Hough representations for lane detection. We subsequently extend its use for semi-supervised training, by noting that the presence of lanes leads to Hough bins with maximal votes.…”