A hybrid model that first finds object instances using either RetinaMask [1] or MaskRCNN [2], whichever has the better mAP on the validation set. For each instance the point-pair voting of [3] is used to recover the pose. For datasets with textured models, symmetries are resolved using the descriptors of [4]. For [3], we use the implementation in MVTec HALCON 20.05.

[1] Fu, C. Y., Shvets, M., & Berg, A. C. (2019). RetinaMask: Learning to predict masks improves state-of-the-art single-shot detection for free. arXiv preprint arXiv:1901.03353. [2] He, K., Gkioxari, G., Dollár, P., & Girshick, R.: Mask r-cnn. ICCV 2017. [3] Drost, B., Ulrich, M., Navab, N., Ilic, S.: Model globally, match locally: Efficient and robust 3d object recognition. CVPR 2010. [4] Lepetit, V., Fua, P.: Keypoint recognition using randomized trees. T-PAMI 2006.