PLACE: Proximity Learning of Articulation and Contact in 3D Environments

Official Pytorch implementation for 2020 3DV paper "PLACE: Proximity Learning of Articulation and Contact in 3D Environments"

[Project] [Paper] [2-min Video] [7-min Video]

<img src="images/teaser.jpg" width = 900 align=middle>

Installation

The code has been tested on Ubuntu 18.04, python 3.6 and CUDA 10.0. Please install the following packages/models following their instructions:

• install SMPL-X, and download SMPL-X models
• install Chamfer Pytorch (source code included in our repo, chamfer_pytorch folder)
• download the VPoser model (source code included in our repo, human_body_prior folder)

Then run pip install -r requirements.txt to install other dependencies. It is noticed that different versions of smplx and VPoser will influece generation results.

Datasets

PROX

PROX dataset can be downloaded here. The following contents are required for this project:

• body_segments the body segments of PROX
• cam2world camera to world transform matrix
• PROXD body mesh of PROX
• sdf scene SDFs of PROX
• scenes the full scene mesh of PROX (optional)
• scenes_downsampled downsampled scene mesh can be downloaded here

Note that you can use either scenes or scenes_downsampled for preprocessing / training / testing. In the paper, we use scenes_downsampled for the whole pipeline.

Matterport3D

We use 7 scenes from Matterport3d for evaluation. Please download scene meshes here and download the computed scene SDFs here Please organize the matterport3D dataset folder as following:

mp3d
├── sdf
|   ├── SCENE_NAME_sdf.npy
|   ├── SCENE_NAME.json
├── SCENE_NAME.ply

• 7 scenes used for evaluation: 17DRP5sb8fy-bedroom, 17DRP5sb8fy-familyroomlounge, 17DRP5sb8fy-livingroom, sKLMLpTHeUy-familyname_0_1, X7HyMhZNoso-livingroom_0_16, zsNo4HB9uLZ-bedroom0_0, zsNo4HB9uLZ-livingroom0_13
• sdf scene SDFs for room meshes
• SCENE_NAME.ply room meshes

Replica

Download Replica dataset here, and organize the dataset folder as following:

replica_v1
├── sdf
|   ├── SCENE_NAME_sdf.npy
|   ├── SCENE_NAME.json
├── SCENE_NAME
|   ├── mesh.ply

• 5 scenes used for evaluation: office_2, hotel_0, room_0, frl_apartment_0, apartment_1

Trained Checkpoints

The trained checkpoints can be downloaded here. Unzip the downloaded checkpoints folder, and put it in the project folder.

Project
├── checkpoints
|   ├── body_dec_last_model.pkl
|   ├── cVAE_last_model.pkl
|   ├── sceneBpsAE_last_model.pkl
|   ├── sceneBpsVertsAE_last_model.pkl
|── ...

Quick Demo

We provide quick demo demo.ipynb to generate random bodies given a scene mesh. Please install all required dependencies and download the checkpoints first.

Preprocessing

Compute the bps encoding and prepare training data with PROX dataset for each scene separately:

python compute_bps_encoding.py --dataset_path PATH/TO/PROX --scene_name SCENE/NAME --split train/test --auge_per_sample 4 --preprocess_file_path PATH/TO/PREPROCESS/RESULTS --smplx_model_path PATH/TO/SMPLX/MODEL

• training scenes: BasementSittingBooth, MPH8, MPH11, MPH112, N0Sofa, N3Library, Werkraum, N3Office, set --split train and --auge_per_sample 4
• testing scenes: MPH1Library, MPH16, N0SittingBooth, N3OpenArea, set --split test and --auge_per_sample 1
• preprocessed data for testing scenes are meant for evaluation of training loss during training

Note each training scene is augmented 4 times (--auge_per_sample 4) and this preprocessing has a heary requirement for CPU memory. If you run the preprocess script for each training scene multiple times (e.x., 2 times, then set --auge_per_sample 2 each time), the preprocessed files can look like:

Project
├── preprocessed_encoding
|   ├── BasementSittingBooth_1_rot_list.npy
|   ├── BasementSittingBooth_1_scene_bps_list.npy
|   ├── ...
|   ├── BasementSittingBooth_2_rot_list.npy
|   ├── BasementSittingBooth_2_scene_bps_list.npy
|   ├── ...
|   ├── ...
|   ├── MPH1Library_rot_list.npy
|   ├── MPH1Library_scene_bps_list.npy
|   ├── ...
|   ├── ...
|── ...

Training

With preprocessed data, train the model (remember to modify train_scene_list in train.py if you run preprocess script each scene multiple times):

python train.py --dataset_path PATH/TO/PROX --preprocess_file_path PATH/TO/PREPROCESS/RESULTS --weight_loss_kl 0.5 --weight_loss_contact 0.01 --start_contact_loss 200

Testing and Optimization

Load trained checkpoints, random generate bodies given a scene mesh and implement the interaction-based optimization:

python test_optimize.py --optimize --dataset prox/mp3d/replica --dataset_path PATH/TO/DATASET --scene_name SCENE/NAME --prox_dataset_path PATH/TO/PROX --smplx_model_path PATH/TO/SMPLX/MODEL --vposer_model_path PATH/TO/VPOSE/MODEL --save_path PATH/TO/SAVE/RESULTS

Evaluation

Evaluate the physical plausibility metrics of optimized results for each scene, set --visualize True to visualize results:

python eval_contact_collision.py --dataset prox/mp3d/replica --dataset_path PATH/TO/DATASET --scene_name SCENE/NAME --optimize_result_dir PATH/TO/OPTIMIZE/RESULTS --smplx_model_path PATH/TO/SMPLX/MODEL --vposer_model_path PATH/TO/VPOSE/MODEL

Evaluate the diversity metrics for optimized results for all scenes of each dataset:

python eval_diversity.py --dataset prox/mp3d/replica --optimize_result_dir PATH/TO/OPTIMIZE/RESULTS --smplx_model_path PATH/TO/SMPLX/MODEL

Citation

When using the code/figures/data/video/etc., please cite our work

@inproceedings{PLACE:3DV:2020,
  title = {{PLACE}: Proximity Learning of Articulation and Contact in {3D} Environments},
  author = {Zhang, Siwei and Zhang, Yan and Ma, Qianli and Black, Michael J. and Tang, Siyu},
  booktitle = {International Conference on 3D Vision (3DV)},
  month = nov,
  year = {2020},
  month_numeric = {11}
}

Acknowledgments & Disclosure

Acknowledgments. We sincerely thank Mohamed Hassan for his scene SDF calculation protocol. This work was partially supported by the German Research Foundation (DFG): SFB 1233, Robust Vision: Inference Principles and Neural Mechanisms, TP XX, project number: 276693517. Qianli Ma acknowledges the support from the Max Planck ETH Center for Learning Systems.

Disclosure. MJB has received research gift funds from Intel, Nvidia, Adobe, Facebook, and Amazon. While MJB is a part-time employee of Amazon, his research was performed solely at MPI. He is also an investor in Meshcapde GmbH.

Relevant Projects

Generating 3D People in Scenes without People (CVPR 2020) ([paper] [code])