Multiview Scene Graph (NeurIPS 2024)

This is the official implementation of

Multiview Scene Graph

Juexiao Zhang, Gao Zhu, Sihang Li, Xinhao Liu, Haorui Song, Xinran Tang, Chen Feng

New York University

[Project page] | [Paper]

Implementations

Requirements

First, setup the environment by running

git clone https://github.com/ai4ce/MSG.git
cd msg
conda create --name msg python=3.11.8
conda activate msg
pip install -r requirements.txt

This requirements.txt contains minimum dependencies estimated by running pipreqs.

Alternatively, to fully replicate the environment you can also run:

git clone https://github.com/ai4ce/MSG.git
cd msg
conda env create -f environment.yml
conda activate msg

Data and weights

MSG data is converted from Apple's ARKitScenes by transforming its 3D annotations to 2D. The converted dataset can be found at this Dataset Hub on Huggingface. We have also kept the code snippets for data convertion in data_preprocess.

To use the data, download and unzip the data to ./data/msg

• TODO: specify the data usage.

mkdir -p data/msg

We also provide pretrained checkpoint of our AoMSG model in the same hub.

To use the checkpoint, download it to ./exp-results/aomsg

• TODO: specify the checkpoint usage

mkdir -p exp-results/aomsg

Inference

To do inference with the pretrained weights, run:

python inference.py --experiment inference

which loads configurations from the file ./configs/experiments/inference.yaml, where the dataset path and the evaluation checkpoint are specified. You can also specify them via arguments which will overwrite the YAML configs. For example:

python inference.py --experiment inference \
--dataset_path PATH/TO/DATASET \
--eval_output_dir PATH/TO/MODEL/CHECKPOINT \
--eval_chkpt CHECKPOINT/FILE

Additional to inference, you can also leverage MSG for topological localization. Please see localization.py for details.

Training

To train the AoMSG model for MSG:

python train.py --experiment aomsg

To train the SepMSG baselines:

python train.py --experiment sepmsg

Please refer to the respective configuration files ./configs/experiments/aomsg.yaml and ./configs/experiments/sepmsg.yaml for the detailed settings.

To resume training of a pretrained checkpoint, set resume=True and specify the resume_path to the checkpoint in the corresponding YAML configuration files.

For evaluation, simply change the script while keep the same experiment configuration, in which eval_output_dir and eval_chkpt are specified.

# evaluate AoMSG
python eval.py --experiment aomsg 
# evaluate SepSMG
python eval.py --experiment sepmsg 
# evaluate SepMSG-direct, which directly use features from froze backbone for MSG
python eval.py --experiment direct 

NOTE:

This release focuses on the implementation of MSG. Object detection dependency is not included. To use detection results instead of groundtruth detection, we can specify detection results in files and give the result_path as is illustrated in ./configs/experiments/aomsg_gdino.yaml where detection results obtained from GroundingDINO is used.

This means you need to run detection separately and save the results to a path. In the data hub we provide the gdino results for convenience. In the future release, we may include a version incorporating online detection.

BibTex

@inproceedings{
zhang2024multiview,
title={Multiview Scene Graph},
author={Juexiao Zhang and Gao Zhu and Sihang Li and Xinhao Liu and Haorui Song and Xinran Tang and Chen Feng},
booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
year={2024},
url={https://openreview.net/forum?id=1ELFGSNBGC}
}