Official implementation of the paper "HUMOF: Human Motion Forecasting in Interactive Social Scenes" (ICLR 2026).

TODO

• Dataset preprocess
  • GTA-IM
  • Humanise
  • HOI-M3
  • HIK
• Train & Eval code
• Pre-trained weights

Environment setup

conda create -n "humof" python=3.9.19 -y
conda activate humof
pip install -r requirements.txt

Dataset preprocess

Humanise dataset

1. Download raw humanise dataset: follow https://github.com/Silverster98/HUMANISE to prepare data

cd datasets/dataset_preprocess/humanise
python preprocess.py \
--humanise_dir=<humanise_dir> \
--scene_dir=<scene_dir> \
--smplx_dir=<smplx_dir>

Where <humanise_dir>, <scene_dir>, and <smplx_dir> refer to the data from Step 1. The expected directory structure is:

<humanise_dir>
├── motions/
│   ├── 000000.pkl
│   ├── 000001.pkl
│   ├── 000002.pkl
│   ├── 000003.pkl
│   ├── 000004.pkl
│   ├── 000005.pkl
│   ├── 000006.pkl
│   ├── 000007.pkl
│   ├── 000008.pkl
│   ├── 000009.pkl
│   ├── 000010.pkl
│   └── ... (19637 more items)
├── annotation.csv


<scene_dir>
├── <scanId>
│   ├── <scanId>_vh_clean_2.ply
│   └── ... 
├── <scanId>
│   ├── <scanId>_vh_clean_2.ply
│   └── ... 
└── ...


<smplx_dir>
└── smplx/
    ├── SMPLX_FEMALE.npz
    ├── SMPLX_MALE.npz
    ├── SMPLX_NEUTRAL.npz
    └── md5sums.txt

HIK dataset

cd datasets/dataset_preprocess/hik
git clone https://github.com/felixbmuller/SAST.git
cd SAST
conda create -n "SAST" python=3.10
conda activate SAST
pip install -r requirements.txt

Download the Humans in Kitchens and unpack its content to ./data/ (note that . is datasets/dataset_preprocess/hik/SAST now), such that ./data/ contains poses/, scenes/, and body_models/.

cd .. # datasets/dataset_preprocess/hik/SAST -> datasets/dataset_preprocess/hik
python preprocess.py hik hik_shortterm.yaml

Train & Eval

Modify the DATASET_name field in conf0.py and run python main.py