Installations

August 17, 2025 · View on GitHub

Environment Setup

cd /path/to/HSMR
git submodule update --init  # fetch thirdparty code

# Use conda to create environment.
conda create -n hsmr python=3.10
conda activate hsmr

# Alternatively, you can use venv.
python3 --version  # ensure 3.8 or higher
python3 -m venv .hsmr_env
source .hsmr_env/bin/activate

# Install Dependencies
pip install -r requirements.txt  # make sure torch version is aligned with $CUDA_HOME
pip install "git+https://github.com/facebookresearch/detectron2.git"
pip install "git+https://github.com/mattloper/chumpy"
pip install -e .

# Install SKEL from https://github.com/MarilynKeller/SKEL
pip install -e thirdparty/SKEL

If you encounter any version conflicts with requirements.txt, we also provide docs/requirements_py3.8.txt with version-annotated dependencies.

We evaluate our approach on multiple benchmarks. Specifically, we use the benchmarks used by HMR2.0, including 2D datasets (COCO, LSP-EXTENDED, PoseTrack) and 3D datasets (3DPW, H36M). Moreover, we evaluate on the MOYO dataset.

Prepare the evaluation datasets used by HMR2.0:
1. Download hmr2_evaluation_data.tar.gz from DropBox or HuggingFace, or you can simply run this:
```
wget -c 'https://huggingface.co/IsshikiHugh/HSMR-data_inputs/resolve/main/hmr2_evaluation_data.tar.gz' \
    -O hmr2_evaluation_data.tar.gz
```
2. Prepare 3DPW, H36M, COCO, LSP-EXTENDED, and PoseTrack datasets. Check the "data structure" bellow to see how to organize the data. Only the images are required.

Prepare the MOYO evaluation dataset:

Download hsmr_evaluation_data.tar.gz from HuggingFace, or you can simply run this:

wget -c 'https://huggingface.co/IsshikiHugh/HSMR-data_inputs/resolve/main/hsmr_evaluation_data.tar.gz' \
    -O hsmr_evaluation_data.tar.gz

Download MoYo images from here.

Add to the code base.

# Prepare meta data.
tar -xzvf hmr2_evaluation_data.tar.gz -C data_inputs/
rm hmr2_evaluation_data.tar.gz
tar -xzvf hsmr_evaluation_data.tar.gz -C data_inputs/
rm hsmr_evaluation_data.tar.gz

# Prepare image data.
mkdir -p data_inputs/datasets
ln -s /path/to/3dpw      data_inputs/datasets/3dpw
ln -s /path/to/h36m      data_inputs/datasets/h36m
ln -s /path/to/coco      data_inputs/datasets/coco
ln -s /path/to/hr-lspet  data_inputs/datasets/hr-lspet
ln -s /path/to/posetrack data_inputs/datasets/posetrack
ln -s /path/to/moyo      data_inputs/datasets/moyo

After that, the data structure should fit:

data_inputs
├── datasets  # The image files.
│   ├── 3dpw
│   │   └── imageFiles/*/*.jpg
│   ├── coco
│   │   └── val2017/*.jpg
│   ├── h36m
│   │   └── images/*.jpg
│   ├── hr-lspet
│   │   └── *.png
│   ├── posetrack
│   │   └─── posetrack2018/posetrack_data/images/*/*/*.jpg
│   └── moyo
│       ├── 220923_yogi_body_hands_03596_Boat_Pose_or_Paripurna_Navasana_-a
│       │   └── YOGI_Cam_*/*.jpg
│       ├── 220923_yogi_body_hands_03596_Boat_Pose_or_Paripurna_Navasana_-b
│       │   └── YOGI_Cam_*/*.jpg
│       └── ...
├── hmr2_evaluation_data  # The packaged labels of standard benchmark datasets.
│   ├── 3dpw_test.npz
│   ├── coco_val.npz
│   ├── h36m_val_p2.npz
│   ├── hr-lspet_train.npz
│   └── posetrack_2018_val.npz
└── hsmr_evaluation_data  # The packaged labels of extra benchmark datasets.
    └── moyo_v2.npz

🏋 Training Datasets

Download training data parts through the commands below (or click the links in the table):

mkdir -p data_inputs/hsmr_training_data
wget "https://www.dropbox.com/scl/fi/tdnyxoufx8u3f7kcgruah/hsmr_training_data.part1.tar.gz?rlkey=92pm05qa8pimjhrpwipu56svn&st=id2dvek4&dl=1" \
    -O data_inputs/hsmr_training_data/hsmr_training_data.part1.tar.gz
wget "https://www.dropbox.com/scl/fi/az8opeka0zyhl6mk0gj6p/hsmr_training_data.part2.tar.gz?rlkey=7iv6t1hl95ok6zuxp2nit29fr&st=5jepxumi&dl=1" \
    -O data_inputs/hsmr_training_data/hsmr_training_data.part2.tar.gz
wget "https://www.dropbox.com/scl/fi/rd9uribnjvyj896cqbj7o/hsmr_training_data.part3.tar.gz?rlkey=b821r5qslbuqivqav8qvu8fee&st=n215nv19&dl=1" \
    -O data_inputs/hsmr_training_data/hsmr_training_data.part3.tar.gz
wget "https://www.dropbox.com/scl/fi/pxm75g95nbkqxg3er8ozd/hsmr_training_data.part4.tar.gz?rlkey=7e5ftkbzre0smbxi2mtukijuk&st=0bje2g3n&dl=1" \
    -O data_inputs/hsmr_training_data/hsmr_training_data.part4.tar.gz
wget "https://www.dropbox.com/scl/fi/3us8qbxra7v01marw52g3/hsmr_training_data.part5.tar.gz?rlkey=737cofc38z9imafk016n4lxmz&st=qowg65xy&dl=1" \
    -O data_inputs/hsmr_training_data/hsmr_training_data.part5.tar.gz
wget "https://www.dropbox.com/scl/fi/2du8j0wll367mxmqk8u37/hsmr_training_data.part6.tar.gz?rlkey=cdz3trhq1eycyko0ahba5ojen&st=ke4a306z&dl=1" \
    -O data_inputs/hsmr_training_data/hsmr_training_data.part6.tar.gz
wget "https://www.dropbox.com/scl/fi/kb2gv7z45d9ha3clof33j/hsmr_training_data.part7.tar.gz?rlkey=7sonxtqqhuctvzfrpecdnbpd1&st=lbv2oh5n&dl=1" \
    -O data_inputs/hsmr_training_data/hsmr_training_data.part7.tar.gz

file name	sha1sum	size
hsmr_training_data.part1.tar.gz	`6675c3f5987186893c4bc5c8616a5ce743ce941d`	30G
hsmr_training_data.part2.tar.gz	`dd62b3e86e8c8ae33436aa5e54d4e9ef27a64503`	41G
hsmr_training_data.part3.tar.gz	`97a98f70488fddff2ac7a4055b4a831039d95ac4`	39G
hsmr_training_data.part4.tar.gz	`5ac1f308fe0c0e96db7c27308292ae03d2fd9d1f`	39G
hsmr_training_data.part5.tar.gz	`b0898c9ef37f417c15589224199948476cfc3abe`	39G
hsmr_training_data.part6.tar.gz	`c7cc0a94f439fd4614b2c97d62767bcc2332f2a2`	37G
hsmr_training_data.part7.tar.gz	`c62b5f96d0168a5ac0eacb772ecc3eb55c9b2756`	39G

Alternatively, you can download the un-packaged webdataset tars from HuggingFace Dataset.

Check the integrity of the downloaded files:
```
sha1sum hsmr_training_data.part*.tar.gz
```

Unzip and add to the codebase.

cd data_inputs/hsmr_training_data
for file in hsmr_training_data.part*.tar.gz; do
    tar -xzvf "$file"
done

Installations

Environment Setup

Data Preparation

Quick Start

Overview for Advanced Setup

Human Body Models

Checkpoints

Datasets