MVP: One-Shot Object Pose Estimation by Matching With Visible Points

We introduce a novel method for one-shot object pose estimation. Recent detector-free one-shot methods have achieved promising results for challenging low-textured objects. The features in a query image are directly matched with all features in an object point cloud reconstructed via Structure-from-Motion (SfM) techniques. Rejecting invisible 3D points, as well as associated features, is performed implicitly using a deep neural network that is trained specifically for feature matching. This tightly-coupled strategy is prone to preserve 3D points that are rarely visible from the query view. In contrast, we propose to prune such erroneous points using the explicit image-point relational graph, which is a lightweight by-product of the SfM reconstruction. By injecting the graph-based pruning into stacked feature transformers, our method is able to obtain high quality 2D-3D correspondences through matching with visible points in an early stage. The experiments demonstrate that our method outperforms state-of-the-art model-free one-shot methods with faster speed.