Abstract
Our goal is to design architectures that retain the groundbreakingperformance of Convolutional Neural Networks (CNNs) for landmark localizationand at the same time are lightweight, compact and suitable for applicationswith limited computational resources. To this end, we make the followingcontributions: (a) we are the first to study the effect of neural networkbinarization on localization tasks, namely human pose estimation and facealignment. We exhaustively evaluate various design choices, identifyperformance bottlenecks, and more importantly propose multiple orthogonal waysto boost performance. (b) Based on our analysis, we propose a novelhierarchical, parallel and multi-scale residual architecture that yields largeperformance improvement over the standard bottleneck block while having thesame number of parameters, thus bridging the gap between the original networkand its binarized counterpart. (c) We perform a large number of ablationstudies that shed light on the properties and the performance of the proposedblock. (d) We present results for experiments on the most challenging datasetsfor human pose estimation and face alignment, reporting in many casesstate-of-the-art performance. (e) We further provide additional results for theproblem of facial part segmentation. Code can be downloaded fromhttps://www.adrianbulat.com/binary-cnn-landmark