lithiumice/models_hub

Model Repository Documentation

Repository Structure Overview

The repository is organized into eight main directories, each serving a specific purpose in the pipeline:

Meta Data (1_meta_data)

Contains AMASS dataset metadata specifically focused on copycat and occlusion information, essential for motion capture applications.

MediaPipe Models (2_mediapipe_ckpts)

Houses MediaPipe's specialized models for facial landmarks and hand tracking, providing fundamental capabilities for human pose estimation.

4DHumans Framework (3_4DHumans)

Incorporates the SMPL (Skinned Multi-Person Linear Model) framework along with training artifacts. The directory includes model parameters, joint regressors, and HMR2 (Human Mesh Recovery) training checkpoints with corresponding configuration files.

SMPLhub (4_SMPLhub)

Serves as a comprehensive collection of human body models, including:

• MANO (hand model) parameters for both left and right hands
• SMPL models in various formats (NPZ and PKL) for male, female, and neutral body types
• SMPLH (SMPL with detailed hand articulation)
• SMPLX (extended SMPL model with face and hand expressions)

Additional Components

• S3FD (5_S3FD): Contains face detection model weights
• SyncNet (6_SyncNet): Includes audio-visual synchronization model
• SGHM (7_SGHM): Houses ResNet50-based model weights
• KonIQ (8_koniq): Contains pre-trained weights for image quality assessment

├── 1_meta_data
│   └── amass_copycat_occlusion_v3.pkl
├── 2_mediapipe_ckpts
│   ├── face_landmarker.task
│   └── hand_landmarker.task
├── 3_4DHumans
│   ├── data
│   │   ├── smpl
│   │   │   └── SMPL_NEUTRAL.pkl
│   │   ├── smpl_mean_params.npz
│   │   └── SMPL_to_J19.pkl
│   └── logs
│       └── train
│           └── multiruns
│               └── hmr2
│                   └── 0
│                       ├── checkpoints
│                       │   └── epoch=35-step=1000000.ckpt
│                       ├── dataset_config.yaml
│                       └── model_config.yaml
├── 4_SMPLhub
│   ├── 0_misc_files
│   │   └── J_regressor_coco.npy
│   ├── MANO
│   │   └── pkl
│   │       ├── MANO_LEFT.pkl
│   │       ├── mano_mean_params.npz
│   │       └── MANO_RIGHT.pkl
│   ├── SMPL
│   │   ├── basicmodel_X_lbs_10_207_0_v1.1.0_pkl
│   │   │   ├── basicmodel_f_lbs_10_207_0_v1.1.0.pkl
│   │   │   ├── basicmodel_m_lbs_10_207_0_v1.1.0.pkl
│   │   │   └── basicmodel_neutral_lbs_10_207_0_v1.1.0.pkl
│   │   ├── X_model_npz
│   │   │   ├── SMPL_F_model.npz
│   │   │   ├── SMPL_M_model.npz
│   │   │   └── SMPL_N_model.npz
│   │   └── X_pkl
│   │       ├── SMPL_FEMALE.pkl
│   │       ├── SMPL_MALE.pkl
│   │       └── SMPL_NEUTRAL.pkl
│   ├── SMPLH
│   │   ├── X_npz
│   │   │   ├── SMPLH_FEMALE.npz
│   │   │   ├── SMPLH_MALE.npz
│   │   │   └── SMPLH_NEUTRAL.npz
│   │   └── X_pkl
│   │       ├── SMPLH_female.pkl
│   │       ├── SMPLH_male.pkl
│   │       └── SMPLH_NEUTRAL.pkl
│   └── SMPLX
│       ├── mod
│       │   └── SMPLX_MALE_shape2019_exp2020.npz
│       └── X_npz
│           ├── SMPLX_FEMALE.npz
│           ├── SMPLX_MALE.npz
│           └── SMPLX_NEUTRAL.npz
├── 5_S3FD
│   └── sfd_face.pth
├── 6_SyncNet
│   └── syncnet_v2.model
├── 7_SGHM
│   └── SGHM-ResNet50.pth
└── 8_koniq
    └── koniq_pretrained.pkl

Create New Model Repo

Update LFS files

git lfs track "*.gif"
git lfs track "*.jpg"
git lfs track "*.png"

# 4. 使用 git lfs migrate 命令转换现有文件
# 这会将已经提交的文件转换为 LFS 对象
git lfs migrate import --include="*.gif,*.jpg,*.png" --everything

# 5. 强制推送更新后的历史
git push --force origin main

Add new repo

git add .
git commit -m "init"
git push