SO-Pose

This repository contains codes of ICCV2021 paper: SO-Pose: Exploiting Self-Occlusion for Direct 6D Pose Estimation This paper is basically an incremental work to paper: GDR-Net: Geometry-Guided Direct Regression Network for Monocular 6D Object Pose Estimation. We analyze and leverage self-occlusion in 6D pose estimation.

Datasets

The code is based on the released code of GDR-Net in this git (The code of GDR-Net is already included) The struture of the datasets is the same.

Since we need ground truth 2D-3D matching and self-occlusion results, we provide generation methods in .gdrn_selfocc_modeling/tools. Please refer to generate_*.py. Note that public renderers (e.g. EGL, GLUMPY) may introduce noise in rendering, the inherent relations between P (2D-3D matching) and Q (self-occlusion) are not guaranteed. So if you use a renderer for efficiency, please make sure that P and Q lie on the same line.

Training and Testing

Please directly run ./gdrn_selfocc_modeling/main_gdrn.py for training and testing.

Important parameters include

config-file : the path to the configuration file.

resume: if 'True', continue the training process from the last checkpoint.

eval-only: if 'True', directly evalute the model.

Trained Models

The trained models can be downloaded here. PLease unzip the trained models in the directory specified in the configuration file. An example output of the evaluation on LMO is provided.

If you find the code useful, please cite the following papers:

[1]Wang, G., Manhardt, F., Tombari, F., & Ji, X. (2021). GDR-Net: Geometry-Guided Direct Regression Network for Monocular 6D Object Pose Estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 16611-16621).

[2]Di, Y., Manhardt, F., Wang, G., Ji, X., Navab, N., & Tombari, F. (2021). SO-Pose: Exploiting Self-Occlusion for Direct 6D Pose Estimation. arXiv preprint arXiv:2108.08367.