Learning Local Descriptor for Comparing Renders with Real Images

We present a method to train deep-network-based feature descriptor calculate discriminative local descriptions from renders and corresponding real images with similar geometry. are interested in using such descriptors for automatic industrial visual inspection whereby the camera has been coarsely localized respect relatively large mechanical assembly presence of certain components needs be checked compared reference computer-aided design model (CAD). aim perform task by comparing image render textureless 3D CAD learned descriptors. The was trained capture geometric features while staying invariant domain. Patch pairs training were extracted semisupervised manner small data set 100 that manually finely registered starting coarse localization camera. Due size set, network initialized weights classification on ImageNet. A two-step is proposed addressing problem domain adaptation. first, “bootstrapping”, obtain good initial second step, triplet-loss training, provides extracting comparable l2 distance. tested through two approaches: finding correspondences between nearest neighbor matching transforming into Bag Visual Words (BoVW) histograms. observed learning robust cross-domain feasible, even might interest CAD-based assemblies, related applications as tracking or augmented reality. To best our knowledge, this first work reports images.