MUGL

The official pytorch implementation of MUGL: Large Scale Multi Person Conditional Action Generation with Locomotion, WACV 2022. Please visit our webpage for more details.

Table of contents:

1. Description
2. How to use

Description

We introduce MUGL, a novel deep neural model for large-scale, diverse generation of single and multi-person pose-based action sequences with locomotion. Our controllable approach enables variable-length generations customizable by action category, across more than 100 categories. To enable intra/inter-category diversity, we model the latent generative space using a Conditional Gaussian Mixture Variational Autoencoder. To enable realistic generation of actions involving locomotion, we decouple local pose and global trajectory components of the action sequence. We incorporate duration-aware feature representations to enable variable-length sequence generation. We use a hybrid pose sequence representation with 3D pose sequences sourced from videos and 3D Kinect-based sequences of NTU-RGBD-120. To enable principled comparison of generation quality, we employ suitably modified strong baselines during evaluation. Although smaller and simpler compared to baselines, MUGL provides better quality generations, paving the way for practical and controllable large-scale human action generation.

Quantitative comparison

How to use

Requirements

$ pip install -r requirements.txt 

Data Preprocessing

We use NTU-VIBE dataset for all our experiments. To extract NTU-VIBE dataset, VIBE pose estimation used on the entire NTU RGBD-120 dataset. But VIBE extracts only the local pose of the sequence i.e. human pose with respect to the root joint. For Extracting position of human pose at the global trajectory we use NTU-120 kinect dataset. The whole process of data preprocessing is performed in 7 steps.

Step 1: Extract NTU RGBD-120 skeleton using VIBE pose estimation, and place it in ./dataset/NTU-VIBE directory. Then run the following,

$ mkdir -p dataset/NTU-VIBE/
$ cd data_preprocessing/
$ python rotation.py 

Step 2 Run the following for creating data for 2-person interaction classes.

$ python make_dataset2persons.py 

step 3 In the step we extract skeleton data from NTU-120 kinect dataset. Download NTU-120 kinect data and place the skeleton files in ./dataset/nturgbd_skeletons_s001_to_s017 directory and run the following,

$ python txt2npy.py

Step 4 In this step we extract the root trajectory from the NTU Kinect dataset and map with NTU-VIBE dataset.

$ python map_root.py

Step 5

$ python preprocessing2person.py

Step 6

$ python get_seq_len.py

Step 7

$ python make_subset_2_person.py

Training

You can skip this step if you want to use our pretrained model. To train MUGL from scratch run pose.py. Details of MUGL model can be found in model/. For training MUGL run the following,

$ python pose.py

Visualization

If you want to play with pretrained MUGL model please see visualization.ipynb. Further, if you want to generate SMPL-X mesh from the skeleton please refer to visualize_mesh.ipynb.

Evaluation

Implementation of evaluation metrics can be found in the metrics directory.

Bibtex

Please cite our paper if you end up using it for your own research:

@InProceedings{MUGL,
title = {MUGL: Large Scale Multi Person Conditional Action Generation with Locomotion },
author = {Debtanu Gupta, Shubh Maheshwari, Ravi Kiran Sarvadevabhatla},
booktitle = {WACV},
year = {2022},
}