RoHM

Robust Human Motion Reconstruction via Diffusion

Project Page | Paper

RoHM is a novel diffusion-based motion model that, conditioned on noisy and occluded input data, reconstructs complete, plausible motions in consistent global coordinates. -- we decompose it into two sub-tasks and learn two models, one for global trajectory and one for local motion. To capture the correlations between the two, we then introduce a novel conditioning module, combining it with an iterative inference scheme.

Installation

Creating a clean conda environment and install all dependencies by:

conda env create -f environment.yml

After the installation is complete, activate the conda environment by:

conda activate rohm

Data preparation

AMASS

• Download the SMPL-X neutral annotations from AMASS dataset, and unzip the files.
• To preprocess the raw AMASS data into the format for RoHM, run the following script for each subset, where dataset_name indicates the name of each subset. It will save the processed AMASS data to datasets/AMASS_smplx_preprocessed.

python preprocessing_amass.py --dataset_name=SUBSET_NAME --amass_root=PATH/TO/AMASS --save_root=datasets/AMASS_smplx_preprocessed

PROX

Download the following contents for PROX dataset:

• cam2world, calibration and recordings from official PROX dataset
• keypoints_openpose and mask_joint from here
• and organize the contents as below:

PROX
├── cam2world
├── calibration
├── recordings
├── keypoints_openpose
├── mask_joint

EgoBody

Download the following contents for EgoBody dataset:

• kinect_color, data_splits.csv, calibrations, kinect_cam_params, smplx_camera_wearer_*, smplx_interactee_* from the official EgoBody dataset
• keypoints_cleaned, mask_joint and egobody_rohm_info.csv from here
• and organize the contents as below:

EgoBody
├── kinect_color
├── data_splits.csv
├── smplx_camera_wearer_train
├── smplx_camera_wearer_test
├── smplx_camera_wearer_val
├── smplx_interactee_train
├── smplx_interactee_test
├── smplx_interactee_val
├── calibrations
├── kinect_cam_params
├── keypoints_cleaned
├── mask_joint
├── egobody_rohm_info.csv

egobody_rohm_info.csv includes information of recordings from EgoBody that we used for evaluation of RoHM.

SMPL-X body model

Download SMPL-X body model from here. Note that the latest version is 1.1 while we use 1.0 in the implementation.

Download smplx vertices segmentation smplx_vert_segmentation.json from here.

Other data (checkpoints, results, etc.)

Download the model checkpoints from here. Download other processed/saved data from here and unzip, including:

• init_motions, initialized motion sequences (RoHM input) on PROX and EgoBody
• test_results_release, reconstructed motion sequences (RoHM output) on PROX and EgoBody
• eval_noise_smplx, pre-computed motion noise for RoHM evaluation on AMASS

Organize all downloaded data as below:

RoHM
├── data
│   ├── body_models
│   │   ├── smplx_model
│   │   │   ├── smplx
│   ├── checkpoints
│   ├── eval_noise_smplx
│   ├── init_motions
│   ├── test_results_release
│   ├── smplx_vert_segmentation.json
├── datasets
│   ├── AMASS_smplx_preprocessed
│   ├── PROX
│   ├── EgoBody

Training

RoHM is trained on AMASS dataset.

TrajNet Training

Train the vanilla TrajNet with a curriculum training scheme for three stages, with increasing noise ratios:

python train_trajnet.py --config=cfg_files/train_cfg/trajnet_train_vanilla_stage1.yaml 
python train_trajnet.py --config=cfg_files/train_cfg/trajnet_train_vanilla_stage2.yaml --pretrained_model_path=PATH/TO/MODEL
python train_trajnet.py --config=cfg_files/train_cfg/trajnet_train_vanilla_stage3.yaml --pretrained_model_path=PATH/TO/MODEL

For stage 2 and 3, set pretrained_model_path to the trained checkpoint from the previous stage. To obtain the reported checkpoint, we train for 800k/400k/450k steps for stage 1/2/3, respectively.

TrajNet fine-tuning with TrajControl:

python train_trajnet.py --config=cfg_files/train_cfg/trajnet_ft_trajcontrol.yaml --pretrained_backbone_path=PATH/TO/MODEL

Set pretrained_backbone_path to the pre-trained checkpoint of vanilla TrajNet, and we train for 400k to obtain the reported checkpoint.

PoseNet training

Train PoseNet with a curriculum training scheme for two stages, with increasing noise ratios:

python train_posenet.py --config=cfg_files/train_cfg/posenet_train_stage1.yaml
python train_posenet.py --config=cfg_files/train_cfg/posenet_train_stage2.yaml --pretrained_model_path=PATH/TO/MODEL

For stage 2, set pretrained_model_path to the trained checkpoint from the previous stage. To obtain the reported checkpoint, we train for 300k/200k steps for stage 1/2, respectively.

Test and evaluate on AMASS

Test on AMASS

Test on AMASS with different configurations (corresponds to Tab.1 in the paper) and save reconstructed results to test_results/results_amass_full: Note that running the given configurations with the same random seed cannot guarantee exactly the same number across different machines, however the stochasticity is quite small.

• Input noise level 3, and mask 10% frames out (masking out both trajectory and local body pose):

python test_amass_full.py --config=cfg_files/test_cfg/amass_occ_0.1_noise_3.yaml

• Input noise level 3, and mask out lower body joints:

python test_amass_full.py --config=cfg_files/test_cfg/amass_occ_leg_noise_3.yaml

• Input noise level 5, and mask out lower body joints:

python test_amass_full.py --config=cfg_files/test_cfg/amass_occ_leg_noise_5.yaml

• Input noise level 7, and mask out lower body joints:

python test_amass_full.py --config=cfg_files/test_cfg/amass_occ_leg_noise_7.yaml

Evaluate on AMASS

Calculate the evaluation metrics and visualize/render on reconstructed results on AMASS.

• Input noise level 3, and mask 10% frames out (masking out both trajectory and local pose):

python eval_amass_full.py --config=cfg_files/eval_cfg/amass_occ_0.1_noise_3.yaml --saved_data_path=PATH/TO/TEST/RESULTS

• Input noise level 3, and mask out lower body joints

python eval_amass_full.py --config=cfg_files/eval_cfg/amass_occ_leg_noise_3.yaml --saved_data_path=PATH/TO/TEST/RESULTS

• Input noise level 5, and mask out lower body joints

python eval_amass_full.py --config=cfg_files/eval_cfg/amass_occ_leg_noise_5.yaml --saved_data_path=PATH/TO/TEST/RESULTS

• Input noise level 7, and mask out lower body joints

python eval_amass_full.py --config=cfg_files/eval_cfg/amass_occ_leg_noise_7.yaml --saved_data_path=PATH/TO/TEST/RESULTS

Other flags for visualization and rendering:

• --visualize=True: visualize input/output/GT motions with open3d (with both skeletons and body meshes)
• --render=True: render the input/output/GT motions with pyrender and save rendered results to --render_save_path

Test and evaluate on PROX/EgoBody

Correponds to the experiment setups in Tab.2 and Tab.3 in the paper.

Initialization

To obtain the initial (noisy and partially visible) motions on PROX, we use the following options:

• For RGB-based reconstruction on PROX, we obtain the initial body pose from CLIFF, body shape from PIXIE, and global translation / orientation from MeTRAbs.
• For RGBD-based reconstruction on PROX, we obtain the initial motion from per-frame optimization by adapted code from LEMO.
• For RGB-based reconstruction on EgoBody, we obtain the intial motion from VPoser-t using the code from HuMoR.

We provide our preprocessed initial motion sequence in the folder data/init_motions, and the final output motion sequences from RoHM in the folder data/test_results_release for your reference.

Note that for the following scripts, the intial motions should have z-axis up for PROX, and y-axis up for EgoBody.

Test on PROX/EgoBody

• Test on PROX with RGB-D input (initization sequeces obtained by per-frame optimization), and results will be saved to test_results/results_prox_rgbd:

python test_prox_egobody.py --config=cfg_files/test_cfg/prox_rgbd.yaml --recording_name=RECORDING_NAME

• Test on PROX with RGB input (initization sequeces obtained by regressors), and results will be saved to test_results/results_prox_rgb:

python test_prox_egobody.py --config=cfg_files/test_cfg/prox_rgb.yaml --recording_name=RECORDING_NAME

• Test on EgoBody with RGB input (initization sequeces obtained by VPoser-t as in HuMoR), and results will be saved to test_results/results_egobody_rgb:

python test_prox_egobody.py --config=cfg_files/test_cfg/egobody_rgb.yaml --recording_name=RECORDING_NAME

Evaluate on PROX/EgoBody

Calculate the evaluation metrics and visualize/render on reconstructed results on PROX/EgoBody.

• Evaluate on PROX with RGB-D input:

python eval_prox_egobody.py --config=cfg_files/eval_cfg/prox_rgbd.yaml --saved_data_dir=PATH/TO/TEST/RESULTS --recording_name=RECORDING_NAME

• Evaluate on PROX with RGB input:

python eval_prox_egobody.py --config=cfg_files/eval_cfg/prox_rgb.yaml --saved_data_dir=PATH/TO/TEST/RESULTS --recording_name=RECORDING_NAME

• Evaluate on EgoBody with RGB input:

python eval_prox_egobody.py --config=cfg_files/eval_cfg/egobody_rgb.yaml --saved_data_dir=PATH/TO/TEST/RESULTS --recording_name=RECORDING_NAME

Note: recording_name can be set to:

• sequence recording name: then evaluation is done over this particular sequence.
• 'all': the evaluation is done over all sequences in the subset (used to report numbers in the paper).

Other flags for visualization and rendering:

• --visualize=True: visualize input/output/GT motions with open3d
  • --vis_option=mesh: visualize body
  • --vis_option=skeleton: visualize skeleton
• --render=True: render the input/output/GT motions with pyrender and save rendered results to --render_save_path

Customized Input

If you want to run RoHM on your customized input:

• Step 1: prepare the initial SMPL-X sequences following the data format as in data/init_motions
• Step 2: prepare the joint occlusion mask following the data format as in datasets/PROX/mask_joint
  • If you have the 3D scene mesh, render a depth map from the camera view for the 3D scene, and identify if the 3D joint is occluded by comparing the depth values (we use utils/get_occlusion_mask.py to obtain occlusion masks on PROX dataset)
  • If you do not have the 3D scene mesh, you can use confidence scores from OpenPose or other 2D body detection methods and set jonits with low confidence as occluded
• Step 3: Customized canonicalization depending on the coordinate system:
  • The current implementation enables canonicalization for inital sequences with y (EgoBody), or z (PROX/AMASS) axis up, with the canicalized sequences always with z axis up
  • If your input initial sequences do not follow this, you need to firstly perform proper transformation to obtain sequences with z/y axis up

License

The majority of RoHM is licensed under CC-BY-NC (including the code, released checkpoints, released dataset for initialized / final motion sequences), however portions of the project are available under separate license terms:

• Trimesh, Guided Diffusion, and MDM are licensed under the MIT license;
• konia is licensed under Apache License.

Citation

If you find our work useful in your research, please consider citing:

@inproceedings{zhang2024rohm,
   title={RoHM: Robust Human Motion Reconstruction via Diffusion},
   author={Zhang, Siwei and Bhatnagar, Bharat Lal and Xu, Yuanlu and Winkler, Alexander and Kadlecek, Petr and Tang, Siyu and Bogo, Federica},
   booktitle={CVPR},
   year={2024}
 }