In this letter, we introduce PolarMesh as a star-convex approximation of a 3D object based on spherical projection and can be applied to monocular object pose and shape estimation. The proposed PolarMesh can be stored in a discrete 2D map that allows a trivial conversion between it and the object surface, as the adjacent map elements form faces of the 3D model. The fixed-grid 2D nature of this object surface approximation ensures that it can be estimated from an image patch of the object with an auto-encoder neural network. Moreover, the PolarMesh not only encodes the object shape but also its orientation. Consequently, it is used for accurate estimation of object orientation and can be extended to the full 6D pose estimation using a projection-based method for translation. An exhaustive quantitative evaluation on LineMod, LineMod-Occlusion and HomebrewedDB shows the competitive performance of the proposed method compared to other monocular RGB pose estimation pipelines. Furthermore, we tested the generalization capabilities of the representation on the recent NOCS dataset for category-level pose estimation and achieved very compelling results in terms of predicted rotations.

Precise annotation of 6-D poses in real data is intricate and time-consuming, however, an essential requirement to train pose estimation pipelines. We propose a way for scalable, end-to-end 6-D pose regression with weak supervision to avoid this problem. Our method requires neither 3-D models nor 6-D object poses as ground truth. Instead, we use 2-D bounding boxes and object sizes as the only labels and constrain the problem with multiple images of known relative poses during training. A novel Rotated-IoU loss brings together a pose prediction from an image with labeled 2-D bounding boxes of the corresponding object in other views. Our rotation estimation combines an initial coarse pose classification with an offset regression using a continuous rotation parametrization that allows for direct pose estimation. At test time, the model still uses only a single image to predict a 6-D pose. We observe that multi-view constraints and our rotation representation used during training lead to better learning of 6-D pose embeddings in comparison to fully supervised methods. Experiments on several datasets show that the proposed method is capable of predicting poses of good quality, in spite being trained with only weak labels. Direct pose regression without the need for a consecutive refinement stage thereby ensures real-time performance.