Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

Body Meshes as Points

Code repository for the paper:
Body Meshes as Points
Jianfeng Zhang, Dongdong Yu, Jun Hao Liew, Xuecheng Nie, Jiashi Feng
CVPR 2021
[paper] [project page (coming soon)]


Overview of the proposed method:
alt text

Installation instructions

This codebase was adopted from early version of mmdetection, mmcv and multiperson. Users of this repo are highly recommended to read the readme of mmcv and mmdetection and multiperson before using this code.

To install relevant envs:

conda env create -f environment.yml
cd mmcv
python3 install
cd ../mmdetection
python develop
cd ../neural_renderer
python3 install
cd ../sdf
python3 install

Prepare datasets

Please follow multiperson for the preparation of the dataset files.

Run training code

Please make sure you have prepared all datasets before running our training script. The training of our model would take two phases, pretrain -> fine tuning. We prepared two configuration files under mmdetection/configs/bmp/. To train our model from scratch:

cd mmdetection
# Phase 1: Pretraining
python3 tools/ configs/bmp/ --create_dummy
while true:
    python3 tools/ configs/bmp/
# We could move to next phase after training for 200 epoches

# Phase 2: Fine-tuning
python3 tools/ configs/bmp/ --load_pretrain ./work_dirs/pretrain/latest.pth
while true:
    python3 tools/ configs/bmp/ 
# It could be done after 50 epoches of training

Run evaluation code

You could use the trained model to evaluate on Panoptic, MuPoTS-3D, Human3.6M and 3DPW.

Example usage:

cd mmdetection
bash scripts/

Run demo code

You could use the trained model for image inference by running:

cd mmdetection
bash scripts/

Qualitative results:
alt text


If you find this code useful for your research, please consider citing the following paper:

  title       = {Body Meshes as Points},
  author      = {Zhang, Jianfeng and Yu, Dongdong and Liew, Jun Hao and Nie, Xuecheng and Feng, Jiashi},
  booktitle   = {CVPR},
  year        = {2021}


This code uses (mmcv, mmdetection) as backbone, and borrows several code from multiperson, SPIN and VIBE. We gratefully appreciate the impact these libraries had on our work. If you use our code, please consider citing the original papers as well.