This is the official implementation of the ICLR 2025 paper, "Learning to Generate Diverse Pedestrian Movements from Web Videos with Noisy Labels", including the preprocessing of the CityWalkers dataset and the PedGen model code release.
Zhizheng Liu, Joe Lin, Wayne Wu, Bolei Zhou
University of California, Los Angeles
Setup the repo:
git clone --recursive [email protected]:genforce/PedGen.git
cd PedGen
conda env create -f env.yaml -n pedgen
conda activate pedgen
pip install -e .
Download checkpoints at this link in ckpts folder:
pedgen_no_context.ckpt, PedGen model without context factors.pedgen_with_context.ckpt, PedGen model with all context factors (scene, human, goal).
Run Demo:
python scripts/demo.py
Feel free to try different context factors to generate diverse movements.
Please check preprocess.md for details.
We use lightning-cli to do train/eval our model. To train on CityWalkers and reproduce our results run:
python scripts/main.py fit -c cfgs/pedgen_with_context.yaml --data.data_root $DATA_ROOT
, where $DATA_ROOT is the root of the preprocessed CityWalkers dataset. Additional information about CARLA evaluation can be found here.
We would like to thank the following projects for inspiring our work and open-sourcing their implementations: WHAM, SLAHMR, MDM, HumanMAC, TRUMANS, ZoeDepth, SegFormer, SLOPER4D.
For any questions or discussions, please contact Zhizheng Liu.
If our work is helpful to your research, please cite the following:
@inproceedings{liu2025learning,
title={Learning to Generate Diverse Pedestrian Movements from Web Videos with Noisy Labels},
author={Liu, Zhizheng and Lin, Joe and Wu, Wayne and Zhou, Bolei},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025}
}