Home

Awesome

Projective Manifold Gradient Layer for Deep Rotation Regression (CVPR 2022)

Introduction

This project is based on our paper Projective Manifold Gradient Layer for Deep Rotation Regression. For more information, you can also visit our project page.

Structure

The implementation of our Regularized Projective Manifold Gradient(RPMG) Layer for rotation regression is in utils/rpmg.py. Since our RPMG layer is a plug-in module which needs to be applied to other tasks, most of our codes are revised from other public repositories on github.

To run each experiment, please see the README.md in each corresponding subfolder.

1. ModelNet_PC

3D object pose estimation from ModelNet40 point clouds developed based on RotationContinuity.

<span class="center"><img src="imgs/PC_Training_curve_airplane.png" width="45%"> <img src="imgs/PC_Percentile_airplane.png" width="45%"></span>

Left: median test error of airplane in different iterations during training. Right: test error percentiles of airplane after training completes. See the same color for comparison of w/ and w/o our RPMG layer.

For more results, please see Exp. 5.1&5.3 in our main paper and Exp. 5.3 in supplementary material

2. ModelNet_Img

3D object pose estimation from ModelNet10 images developed based on Spherical_Regression.

<span class="center"><img src="imgs/ModelNetimg_training_curve_chair.png" width="45%"> <img src="imgs/ModelNetimg_percentile_chair.png" width="45%"></span>

Left: median test error of chair in different iterations during training. Right: test error percentiles of chair after training completes. See the same color for comparison of w/ and w/o our RPMG layer.

For more results, please see Exp. 5.2 in our main paper.

4. Pascal3D_Img

3D object pose estimation from Pascal3D images developed based on Spherical_Regression.

<span class="center"><img src="imgs/pascal_training_curve_sofa.png" width="45%"> <img src="imgs/pascal_percentile_sofa.png" width="45%"></span>

Left: median test error of sofa in different iterations during training. Right: test error percentiles of sofa after training completes. See the same color for comparison of w/ and w/o our RPMG layer.

For more results, please see Exp. 5.1 in supplementary material

5. poselstm-pytorch

Camera relocalization developed based on poselstm.

Please see 5.2 in supplementary material

Citation

If you find our work useful in your research, please consider citing:

@article{chen2021projective,
  title={Projective Manifold Gradient Layer for Deep Rotation Regression},
  author={Chen, Jiayi and Yin, Yingda and Birdal, Tolga and Chen, Baoquan and Guibas, Leonidas and Wang, He},
  journal={arXiv preprint arXiv:2110.11657},
  year={2021}
}

License

This work and the dataset are licensed under CC BY-NC 4.0.

CC BY-NC 4.0