Skip to content

naver/multi-hmr

Repository files navigation

Multi-HMR: Multi-Person Whole-Body Human Mesh Recovery in a Single Shot

Fabien Baradel*, Matthieu Armando, Salma Galaaoui, Romain Brégier,
Philippe Weinzaepfel, Grégory Rogez, Thomas Lucas*

ECCV'24

* equal contribution

arXiv Blogpost Demo







Multi-HMR illustration 1 Multi-HMR illustration 2
Multi-HMR is a simple yet effective single-shot model for multi-person and expressive human mesh recovery. It takes as input a single RGB image and efficiently performs 3D reconstruction of multiple humans in camera space.

News

  • 2024/07/03: Release of training-evaluation code.
  • 2024/07/01: Multi-HMR is accepted to ECCV'24.
  • 2024/06/17: Multi-HMR won Robin Challenge @CVPR'24: 3D human reconstruction track.
  • 2024/02/22: Release of demo code.

Installation

First, you need to clone the repo.

We recommand to use virtual enviroment for running MultiHMR. Please run the following lines for creating the environment with venv:

python3.9 -m venv .multihmr
source .multihmr/bin/activate
pip install -r requirements.txt

Otherwise you can also create a conda environment.

conda env create -f conda.yaml
conda activate multihmr

The installation has been tested with python3.9 and CUDA 12.1.

Checkpoints will automatically be downloaded to $HOME/models/multiHMR the first time you run the demo code.

Besides these files, you also need to download the SMPLX model. You will need the neutral model for running the demo code. Please go to the corresponding website and register to get access to the downloads section. Download the model and place SMPLX_NEUTRAL.npz in ./models/smplx/.

Run Multi-HMR on images

The following command will run Multi-HMR on all images in the specified --img_folder, and save renderings of the reconstructions in --out_folder. The --model_name flag specifies the model to use. The --extra_views flags additionally renders the side and bev view of the reconstructed scene, --save_mesh saves meshes as in a '.npy' file.

python3.9 demo.py \
    --img_folder example_data \
    --out_folder demo_out \
    --extra_views 1 \
    --model_name multiHMR_896_L

Pre-trained models

We provide multiple pre-trained checkpoints. Here is a list of their associated features. Once downloaded you need to place them into $HOME/models/multiHMR.

modelname training data backbone resolution runtime (ms) PVE-3PDW-test PVE-EHF PVE-BEDLAM-val comment
multiHMR_896_L BEDLAM+AGORA+CUFFS+UBody ViT-L 896x896 126 89.9 42.2 56.7 initial ckpt
multiHMR_672_L BEDLAM+AGORA+CUFFS+UBody ViT-L 672x672 74 94.1 37.0 58.6 longer training
multiHMR_672_B BEDLAM+AGORA+CUFFS+UBody ViT-B 672x672 43 94.0 43.6 67.2 longer training
multiHMR_672_S BEDLAM+AGORA+CUFFS+UBody ViT-S 672x672 29 102.4 49.3 78.9 longer training
multiHMR_1288_L_bedlam BEDLAM(train+val) ViT-L 1288x1288 ? ? ? ckpt used for BEDLAM leaderboard
multiHMR_1288_L_agora BEDLAM(train+val)+AGORA(train+val) ViT-L 1288x1288 ? ? ? ckpt used for AGORA leaderboard

We compute the runtime on GPU V100-32GB.

Training Multi-HMR

We provide code for training Multi-HMR using a single GPU on BEDLAM-training and evaluating it on BEDLAM-validation, EHF and 3DPW-test.

Activate environnement

source .multihmr/bin/activate
export PYTHONPATH=`pwd`

Preprocessing BEDLAM

The first thing that you need to do is to download the BEDLAM dataset (6fps version) and place the files into data/BEDLAM The data structure of the directory should look like this:

data/BEDLAM
      |
      |---validation
                  |
                  |---20221018_1_250_batch01hand_zoom_suburb_b_6fps
                                                              |
                                                              |---png
                                                                  |
                                                                  |---seq_000000
                                                                              |
                                                                              |---seq_000000_0000.png
                                                                              ...
                                                                              |---seq_000000_0235.png
                                                                  ...
                                                                  |---seq_000249
                  ...
                  |---20221019_3-8_250_highbmihand_orbit_stadium_6fps
      |---training
              |
              |---20221010_3_1000_batch01hand_6fps
              ...
              |---20221024_3-10_100_batch01handhair_static_highSchoolGym_30fps
      |---all_npz_12_training
              |
              |---20221010_3_1000_batch01hand_6fps.npz
              ...
              |---20221024_3-10_100_batch01handhair_static_highSchoolGym_30fps.npz
      |---all_npz_12_validation
            |
            |---20221018_1_250_batch01hand_zoom_suburb_b_6fps.npz
            ...
            |---20221019_3-8_250_highbmihand_orbit_stadium_6fps.npz

We need to build the annotation files for the training and validation sets. It may takes around 20 minutes for bulding the pkl files depending on your CPU.

python3.9 datasets/bedlam.py "create_annots(['validation', 'training'])"

You will get two files data/bedlam_validation.pkl and data/bedlam_training.pkl.

Checking annotations

Visualize the annotation of a specific image.

python3.9 datasets/bedlam.py "visualize(split='validation', i=1500)"

It will create a file bedlam_validation_15000.jpg where you can see the RGB image on the left side and the RGB image with meshes overlayed on the right side.

(Optional) Creating jpg files to fast data-loading

BEDLAM is composed of PNG files and loading them could be a bit slow depending our your infrastucture. The following command will generate one jpg file for each png file with maximal resolution of 1280. It may take a while because BEDLAM has more than 300k images. You can run the command lines on some specific subdirectories to speed-up the generation of jpg files. You can chose the target size of your choice.

# Can be slow
python3.9 datasets/bedlam.py "create_jpeg(root_dir='data/BEDLAM', target_size=1280)

# Or parallelize
python3.9 datasets/bedlam.py "create_jpeg(root_dir='data/BEDLAM/validation/20221019_3-8_250_highbmihand_orbit_stadium_6fps', target_size=1280)
...
python3.9 datasets/bedlam.py "create_jpeg(root_dir='data/BEDLAM/training/20221010_3-10_500_batch01hand_zoom_suburb_d_6fps', target_size=1280)

Checking the data-loading time

You can check the quality of your dataloader by running the command above. It will use the png version of BEDLAM.

python3.9 datasets/bedlam.py "dataloader(split='validation', batch_size=16, num_workers=4, extension='png', img_size=1280, n_iter=100)"

Preprocessing additional validation sets

We also provide code for evaluating on EHF and 3DPW. Run the command for bulding the annotation fiel for EHF.

python3.9 datasets/ehf.py "create_annots()"
python3.9 datasets/ehf.py "visualize(i=10)"

And for 3DPW. Please download SMPL-male and SMPL-female models, put them into models/smpl/SMPL_MALE.pkl and models/smpl/SMPL_FEMALE.pkl. And smplx2smpl.pkl is mandatory for moving from SMPLX to SMPL.

python3.9 datasets/threedpw.py "create_annots()"
python3.9 datasets/threedpw.py "visualize(i=1011)"

Training on BEDLAM-train

We provide the command for training on BEDLAM-train at resolution 336 on a single GPU.

# python command
CUDA_VISIBLE_DEVICES=1 python3.9 train.py \
--backbone dinov2_vits14 \
--img_size 336 \
-j 4 \
--batch_size 32 \
-iter 10000 \
--max_iter 500000 \
--name multi-hmr_s_336

To decrease data-loading time use --extension jpg --res 1280

Evaluating BEDLAM-val / EHF-test / 3DPW-test

Above command is for evaluating a pretrained ckpt on validation sets.

CUDA_VISIBLE_DEVICES=0 python3.9 train.py \
--eval_only 1 \
--backbone dinov2_vitl14 \
--img_size 896 \
--val_data EHF THREEDPW BEDLAM \
--val_split test test validation \
--val_subsample 1 20 25 \
--pretrained models/multiHMR/multiHMR_896_L.pt

Either check the log or open the tensorboard for checking the results.

CUFFS dataset

The Close-Up Frames of Full-Body Subjects dataset, containing humans close to the camera with diverse hand poses is available here(LICENSE). More information about how to use it will be given soon, stay tuned.

License

The code is distributed under the CC BY-NC-SA 4.0 License.
See Multi-HMR LICENSE, Checkpoint LICENSE and Example Data LICENSE for more information.

Citing

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

@inproceedings{multi-hmr2024,
    title={Multi-HMR: Multi-Person Whole-Body Human Mesh Recovery in a Single Shot},
    author={Baradel*, Fabien and 
            Armando, Matthieu and 
            Galaaoui, Salma and 
            Br{\'e}gier, Romain and 
            Weinzaepfel, Philippe and 
            Rogez, Gr{\'e}gory and
            Lucas*, Thomas
            },
    booktitle={ECCV},
    year={2024}
}