3D Object Detection and Pose Estimation of Unseen Objects in Color Images with Local Surface Embeddings

We present an approach for detecting and estimating the 3D poses of objects in images that requires only untextured CAD model no training phase new objects. Our combines Deep Learning geometry: It relies on embedding local geometry to match models input images. For points at surface objects, this can be computed directly from model; image locations, we learn predict it itself. This establishes correspondences between 2D locations However, many these are ambiguous as may have similar geometries. show use Mask-RCNN a class-agnostic way detect without retraining thus drastically limit number possible correspondences. then robustly estimate pose discriminative using RANSAC-like algorithm. demonstrate performance T-LESS dataset, by small testing other experiments our method is par or better than previous methods.