Part-aware Measurement for Robust Multi-View Multi-Human 3D Pose Estimation and Tracking

Part-Aware Measurement for Robust Multi-View Multi-Human 3D Pose Estimation and Tracking
Hau Chu, Jia-Hong Lee, Yao-Chih Lee, Ching-Hsien Hsu, Jia-Da Li, Chu-Song Chen
2021 CVPR B-AMFG Workshop

Note: It's a project of AI^2 Lab. The code will be updated in here while there is a new version.

Installation

• Python 3.6+

• Cuda 9.0

• Cudnn 7

• gcc 5 & g++ 5 (for Ubuntu 18.04)

$ sudo apt install gcc-5 g++-5
$ sudo ln -s /usr/bin/gcc-6 /usr/local/bin/gcc
$ sudo ln -s /usr/bin/g++-6 /usr/local/bin/g++

• Conda Env

$ conda create -n venv python=3.6
$ conda activate venv
$ conda install pytorch==1.1.0 torchvision==0.3.0 cudatoolkit=9.0 -c pytorch
$ pip install tensorflow_gpu==1.9.0
$ pip install -r requirements.txt

• Git

$ sudo apt install git

Data preparation

Download datasets:

1. Campus (http://campar.in.tum.de/Chair/MultiHumanPose)
2. Shelf (http://campar.in.tum.de/Chair/MultiHumanPose)
3. CMU Panoptic (https://github.com/CMU-Perceptual-Computing-Lab/panoptic-toolbox)

Dataset's camera_parameter.pickle download

The directory tree should look like below:

${ROOT}
    |-- CatchImage
        |-- CampusSeq1
        |   |-- Camera0
        |   |-- Camera1
        |   |-- Camera2
        |   |-- camera_parameter.pickle
        |   |-- actorsGT.mat
        |-- Shelf
        |   |-- Camera0
        |   |-- ...
        |   |-- Camera4
        |   |-- camera_parameter.pickle
        |   |-- actorsGT.mat
        |-- Panoptic
        |   |-- 160906_pizza1
            |   |-- 00_03 # hdImgs folder of 03 camera
            |   |-- 00_06 # hdImgs folder of 06 camera
            |   |-- ...
            |   |-- camera_parameter.pickle
            |   |-- hdPose_stage1_coco19
            |-- ...
    |-- src

Backend Models

Backend models, which is not our works, are released codes from others. We only did some small modifications to fit the format of our input/output. Put models in {ROOT}/src/backend

1. YOLOv3
   • origin: https://github.com/ZQPei/deep_sort_pytorch
   • modification
2. HRNet
   • origin: https://github.com/stefanopini/simple-HRNet
   • modification

Run Codes

Demo

$cd src
python -W ignore testmodel.py --dataset CampusSeq1 # For Campus
python -W ignore testmodel.py --dataset Shelf # For Shelf
python -W ignore testmodel.py --dataset Panoptic # For Panoptic (sub-dataset can be modified in config)

Evaluation

$cd src
python -W ignore evalmodel.py --dataset CampusSeq1 
python -W ignore evalmodel.py --dataset Shelf

Campus PCP Score

Bone Group	Actor 0	Actor 1	Actor 2	Average
Head	100.00	100.00	100.00	100.00
Torso	100.00	100.00	100.00	100.00
Upper arms	98.98	100.00	100.00	99.66
Lower arms	92.86	68.78	91.30	84.31
Upper legs	100.00	100.00	100.00	100.00
Lower legs	100.00	100.00	100.00	100.00
Total	98.37	93.76	98.26	96.79

Shelf PCP Score

Bone Group	Actor 0	Actor 1	Actor 2	Average
Head	94.98	100.00	91.30	95.43
Torso	100.00	100.00	100.00	100.00
Upper arms	100.00	100.00	96.27	98.76
Lower arms	98.21	77.03	96.27	90.50
Upper legs	100.00	100.00	100.00	100.00
Lower legs	100.00	100.00	100.00	100.00
Total	99.14	95.41	97.64	97.39

Citation

@InProceedings{Chu_2021_CVPR,
    author    = {Chu, Hau and Lee, Jia-Hong and Lee, Yao-Chih and Hsu, Ching-Hsien and Li, Jia-Da and Chen, Chu-Song},
    title     = {Part-Aware Measurement for Robust Multi-View Multi-Human 3D Pose Estimation and Tracking},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2021},
    pages     = {1472-1481}
}