Awesome

Regression-based 3D Pose Estimation for Texture-less Objects

Introduction

In this paper, we propose a novel network structure to estimate 3D pose for texture-less objects. The network contains two parts, the triplet network, to extract expected features from images; and the regression network, to directly regress the 3D pose using the features output by the triplet network.

Citation

If you find this code useful in your research, please consider citing:

@article{liu2019regression,
    Auther = {Liu, Yuanpeng and Zhou, Laishui and Zong, Hua and Gong, Xiaoxi and Wu, Qiaoyun and Liang, Qingxiao and Wang, Jun},
    Title = {Regression-based 3D Pose Estimation for Texture-less Objects},
    Journal = {IEEE Transactions on Multimedia},
    year={2019}
}

Installation

Suggested environment

• Tensorflow >= 1.4.0
• Keras >= 2.1.0
• CUDA >= 8.0

Instruction

1. Download dataset

2. Unzip dataset:

mkdir ./dataset
cd ./dataset
unzip dataset-3dpr.zip

3. Extract features correlated to poses:

python ./FeatureExtraction.py

4. Regress poses using features:

python ./Regression.py

Note

The matching scores are evaluated using sixd_toolkit, thanks to Tomas Hodan.