Multi-Path Learning for Object Pose Estimation Across Domains

Kurzfassung

We introduce a scalable approach for object pose estima-tion trained on simulated RGB views of multiple 3D modelstogether. We learn an encoding of object views that doesnot only describe an implicit orientation of all objects seenduring training, but can also relate views of untrained ob-jects. Our single-encoder-multi-decoder network is trainedusing a technique we denote multi-path learning: Whilethe encoder is shared by all objects, each decoder only re-constructs views of a single object. Consequently, viewsof different instances do not have to be separated in thelatent space and can share common features. The result-ing encoder generalizes well from synthetic to real dataand across various instances, categories, model types anddatasets. We systematically investigate the learned encod-ings, their generalization, and iterative refinement strate-gies on the ModelNet40 and T-LESS dataset. Despite train-ing jointly on multiple objects, our 6D Object Detectionpipeline achieves state-of-the-art results on T-LESS at muchlower runtimes than competing approaches.