smpl-kit

smplkit is a versatile library that simplifies SMPL body model usage with automatic model path search, different body models support, adjustable batch size support, and useful utility functions for data processing and visualization.

The left figure utilizes smplx for creating an SMPL body mesh with random translation, while smplkit achieves the same functionality with a more concise code implementation.

Features

• Automatic model path search in the workspace or home folder.
• Support SMPL, SMPL+H, SMPL-X body models.
• Support adjustable batch size of the SMPL parameters.
• Provide numerous utility functions for SMPL parameters and meshes.
• Provide visualization tools for viewing the results.

Installation

1. Before installing smplkit, please make sure that you have installed pytorch and pytorch3d.

2. To install the smplkit package, you can either install it from PyPI or install it from source.

   • To install from PyPI by using pip:

   pip install smplkit
   

   • Or, clone this repository and install it from source:

   git clone git@github.com:Silverster98/smpl-kit.git
   cd smpl-kit
   pip install .
   

Documentation

Tutorial

0. Model path configuration

smplkit can search the model automatically. However, you can also specify the model path manually.

• If the model_path is not provided as a parameter for SMPL/SMPL+H/SMPL-X layers, the program will automatically search the models in current folder.
  • If ./body_models/ exists, the program will search the models in ./body_models/.
    • You should organize the ./body_models/ as follows:

      |- ./body_models/
      |--- smpl/  # SMPL model
      |----- SMPL_NEUTRAL.pkl
      |--- smplh/ # SMPL+H model
      |----- SMPLH_NEUTRAL.pkl
      |----- SMPLH_MALE.pkl
      |----- SMPLH_FEMALE.pkl
      |--- smplx/ # SMPL-X model
      |----- SMPLX_NEUTRAL.npz
      |----- SMPLX_MALE.npz
      |----- SMPLX_FEMALE.npz
      

  • If ./body_models does not exist, the program will search the models in ~/.body_models/, i.e., the home folder.
    • You should organize the ~/.body_models/ as follows (similar to the above):

      |- ~/.body_models/
      |--- smpl/  # SMPL model
      |--- smplh/ # SMPL+H model
      |--- smplx/ # SMPL-X model
      

  • So, you can either put the models in one workspace for one project or in the home folder for all related projects.
• If the model_path is provided as a parameter for SMPL/SMPL+H/SMPL-X layers, the program will search the models in the given path.
  • If the model_path is file path, the program will load the model from the given file path.
  • If the model_path if folder path, the program will search the models in the given folder path.
    • You should organize the given folder as follows (similar to the above):

      |- ${GIVEN_FOLDER}/
      |--- smpl/  # SMPL model
      |--- smplh/ # SMPL+H model
      |--- smplx/ # SMPL-X model
      

1. Use the SMPLLayer to generate a SMPL body mesh with random translation while keeping other parameters as zero:

import torch
import trimesh
from smplkit import SMPLLayer as SMPL

body_model = SMPL()

transl = torch.rand((2, 3), dtype=torch.float32)
verts = body_model(transl=transl, return_verts=True)
verts = verts.numpy()

mesh = trimesh.Trimesh(vertices=verts[0], faces=body_model.faces)

2. Use the SMPLParam to optimize the SMPL parameters (only tranlation) with a given mesh:

import torch
from smplkit import SMPLLayer as SMPL
from smplkit import SMPLParam
from torch.optim import SGD

bm = SMPL(num_betas=10)

target_transl = torch.tensor([[0, 1, 0]], dtype=torch.float32)
target_verts = bm(transl=target_transl, return_verts=True)

## init param and optimizer
param = SMPLParam(transl=torch.rand(1, 3), requires_grad=[True, False, False, False])
opt = SGD(param.trainable_parameters(), lr=0.1)

for i in range(200):
    opt.zero_grad()
    output = bm(**param._parameters_dict(), return_verts=True)
    
    loss = ((output - target_verts) ** 2).mean()
    
    loss.backward()
    opt.step()

    if (i + 1) % 20 == 0:
        print(f"Optimization Error in Step {i + 1:3d}: {loss.item()}")

API

• SMPLLayer
• SMPLHLayer
• SMPLXLayer
• SMPLParam
• SMPLHParam
• SMPLXParam
• utils
  • matrix_to_parameter
  • compute_orient
  • compute_normal
  • compute_sdf
  • VertexSelector
    • select_vertex
    • contact_vertex (Only support SMPL-X now)
    • key_vertex
  • JointSelector
    • select_joint
  • BodyModel
    • reset
    • run
    • to
• visualize
• constants
  • VERTEX_NUM
  • JOINTS_NUM
  • KEY_VERTEX_IDS
  • CONTACT_PART_NAME
  • CONTACT_VERTEX_IDS (Only support SMPL-X now)
  • JOINTS_NAME
  • SMPLX_LANDMARKS_NAME
  • SKELETON_CHAIN
  • SKELETON_PARENTS

License

This project is licensed under the terms of the MIT license.

Acknowledgement

Some codes are borrowed from SMPL-X. If your use this code, please consider citing the most relevant works.