Skip to content
Code for CVPR2019 PIE_pose_invariant_embedding
Python Shell
Branch: master
Clone or download
Latest commit ec1ac37 Aug 24, 2019
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
ObjectPI Upload dataset Aug 23, 2019
cnn_based first commit Aug 23, 2019
proxy_based first commit Aug 23, 2019
triplet_center_based first commit Aug 23, 2019
ObjectPI.pickle first commit Aug 23, 2019
README.md Update README.md Aug 23, 2019
download_modelnet40.sh first commit Aug 23, 2019
environment.yml add environment.yml Aug 23, 2019
modelnet.pickle first commit Aug 23, 2019

README.md

PIE_pose_invariant_embedding

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.

Usage

  1. Install all required library
conda env create -f environment.yml --name <env_name>
  1. Download ModelNet40 dataset
download_modelnet40.sh
  1. Download pretrained models from https://drive.google.com/drive/folders/1l9VASmcr2oRD0PKKgv222syhsVcpU290?usp=sharing and place the pretrain models according to the folder organization.

  2. Run cnn based methods (cnn, mvcnn, picnn) by:

cd cnn_based
cd <cnn, mvcnn, picnn>
sh run.sh
  1. Run proxy based methods (proxy, mvproxy, piproxy) by:
cd proxy_based
cd <proxy, mvproxy, piproxy>
sh run.sh
  1. Run triplet center based methods (triplet, mvtriplet, pitriplet) by:
cd triplet_center_based
cd <triplet, mvtriplet, pitriplet>
sh run.sh
  1. 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.

Result for pretrained models

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.

MVCNN

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

ObjectPI

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

Citation

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}
}
You can’t perform that action at this time.