ssfootball04 / generative_pose

generative_pose

Code for our ICCV 19 Paper : Monocular 3D Human Pose Estimation by Generation and Ordinal Ranking, available here : https://arxiv.org/abs/1904.01324

Dependencies

• pytorch 1.0.0
• h5py 2.8.0
• matplotlib 3.0.2

Setup

1. Clone this repository.
2. Download the data from Google Drive and unzip it inside the parent directory. It contains preprocessed ground truth 3D coordinates on Human3.6, and 2D pose + Ordinal Relation detections from our 2DPoseNet/OrdinalNet module.

Running the Code

For training MultiPoseNet, run the following command

python main.py --exp [name of your experiment]

For testing run,

python main.py --exp [name of your experiment] --test --numSamples [num of samples to generate] --load [pre-trained model]

Pre-trained model

We provide a pre-trained model on Google Drive. You can reproduce our results by running the test script with this model and generating 200 samples.

Code Layout

This repository closely follows una_dinosauria's Tensorflow repo for their ICCV 17 paper A Simple Yet Effective Baseline for 3D Pose Estimation, and weigq's Pytorch repo for the same.

Citing

If you use this code, please cite our work :

@InProceedings{Sharma_2019_ICCV,
author = {Sharma, Saurabh and Varigonda, Pavan Teja and Bindal, Prashast and Sharma, Abhishek and Jain, Arjun},
title = {Monocular 3D Human Pose Estimation by Generation and Ordinal Ranking},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {October},
year = {2019}
}