PyraPose: Feature Pyramids for Fast and Accurate Object Pose Estimation under Domain Shift by Stefan Thalhammer, Timothy Patten, and Markus Vincze.

My docker-repository provides Dockerfiles satisfying the version requirements

python setup.py build_ext --inplace

1. annotate data to create training dataset using "annotate_BOP.py" in repo data_creation.
   • the 3D bounding boxes used to establish 2D-3D correspondences are hard coded.
   • Synthetic training data can be taken from the BOP-challenge. A good source for object meshes, test and val data.
2. train using "PyraPose/bin/train.py </path/to/training_data>".

PyraPose/bin/evaluate.py </path/to/dataset_val> </path/to/training/model.h5> --convert-model

Data loaders are provided for datasets LineMOD, Occlusion, YCB-video, HomebrewedDB and Tless. No trained models are provided.

• branch "master" uses provides the proposed method in the paper, i.e., PFPN+heads.
• branch "decoder" provides the method when using decoders with skip-connections instead.