Context-PIPs: Persistent Independent Particles Demands Spatial Context Features

This repository is the source code for paper Context-PIPs: Persistent Independent Particles Demands Spatial Context Features, NeurIPS 2023.

Weikang Bian^*, Zhaoyang Huang^*, Xiaoyu Shi, Yitong Dong, Yijin Li, Hongsheng Li ( ^* denotes equal contributions.)

[Paper] [Project Page]

Requirements

conda create --name context_pips python=3.10
conda activate context_pips
conda install pytorch=1.12.0 torchvision=0.13.0 cudatoolkit=11.3 -c pytorch
pip install -r requirements.txt

Data Preparation

To evaluate/train our Context-PIPs, you will need to download the following datasets.

You can create symbolic links to wherever the datasets were downloaded in the data folder.

├── data
    ├── flyingthings
        ├── frames_cleanpass_webp
        ├── object_index
        ├── occluders_al
        ├── optical_flow
        ├── trajs_ad
    ├── HT21
        ├── test
        ├── train

Evaluation

We provide a model for evaluation.

# Evaluate Context-PIPs on FlyingThings++
python test_on_flt.py --init_dir path_to_checkpoint_folder

# Evaluate Context-PIPs on CroHD
# Occluded
python test_on_crohd.py --init_dir path_to_checkpoint_folder
# Visible
python test_on_crohd.py --init_dir path_to_checkpoint_folder --req_occlusion False

Training

Similiar to PIPs, we train our model on the FlyingThings++ dataset:

python train.py \
    --horz_flip=True --vert_flip=True \
    --device_ids=\[0,1,2,3,4,5,6,7\] \
    --exp_name contextpips \
    --B 4 --N 128 --I 6 --lr 3e-4

Acknowledgement

In this project, we use parts of code in:

Thanks to the authors for open sourcing their code.

Citation

@inproceedings{
bian2023contextpips,
title={Context-{PIP}s: Persistent Independent Particles Demands Context Features},
author={Weikang BIAN and Zhaoyang Huang and Xiaoyu Shi and Yitong Dong and Yijin Li and Hongsheng Li},
booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
year={2023},
url={https://openreview.net/forum?id=cnpkzQZaLU}
}