Implementation of the research paper PIEs: Pose Invariant Embeddings (CVPR2019) The result presented in the paper is averaged over 5 rounds, so it might be slightly different from the model provided.
- Install all required library
conda env create -f environment.yml --name <env_name>
- Download ModelNet40 dataset
download_modelnet40.sh
-
Download pretrained models from https://drive.google.com/drive/folders/1l9VASmcr2oRD0PKKgv222syhsVcpU290?usp=sharing and place the pretrain models according to the folder organization.
-
Run cnn based methods (cnn, mvcnn, picnn) by:
cd cnn_based
cd <cnn, mvcnn, picnn>
sh run.sh
- Run proxy based methods (proxy, mvproxy, piproxy) by:
cd proxy_based
cd <proxy, mvproxy, piproxy>
sh run.sh
- Run triplet center based methods (triplet, mvtriplet, pitriplet) by:
cd triplet_center_based
cd <triplet, mvtriplet, pitriplet>
sh run.sh
- Check the result of pretrained models in log folder and log_robustness folder. Log_robustness folder shows the classification results from single view to all the views provided. Read run.sh for more information.
The proposed method is similar to single view based methods when only single view is given and similar to multiview based methods (mv___) when all views are given. The proposed method is more robusted to the number of view given during inference time.
| Methods | Classification | ||
| Single view | All views (12) | ||
| CNN based | cnn | 84.66 | 87.50 |
| mvcnn | 77.75 | 89.75 | |
| picnn | 85.70 | 89.25 | |
| Proxy based | proxy | 85.60 | 88.62 |
| mvproxy | 78.39 | 90.38 | |
| piproxy | 85.49 | 89.25 | |
| Triplet center based | triplet | 85.23 | 88.75 |
| mvtriplet | 76.94 | 89.38 | |
| pitriplet | 83.50 | 89.38 | |
| Methods | Classification | ||
| Single view | All views (12) | ||
| CNN based | cnn | 65.82 | 76.53 |
| mvcnn | 59.44 | 77.55 | |
| picnn | 67.60 | 79.59 | |
| Proxy based | proxy | 69.52 | 79.59 |
| mvproxy | 64.03 | 76.53 | |
| piproxy | 68.62 | 79.59 | |
| Triplet center based | triplet | 70.79 | 77.55 |
| mvtriplet | 63.65 | 78.57 | |
| pitriplet | 69.64 | 75.51 | |
If you mentioned the method in your research, please cite this article:
@InProceedings{Ho_2019_CVPR,
author = {Ho, Chih-Hui and Morgado, Pedro and Persekian, Amir and Vasconcelos, Nuno},
title = {PIEs: Pose Invariant Embeddings},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2019}
}