EmbodiedScan: A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI

Tai Wang*  Xiaohan Mao*  Chenming Zhu*  Runsen Xu  Ruiyuan Lyu  Peisen Li  Xiao Chen 
Wenwei Zhang  Kai Chen  Tianfan Xue  Xihui Liu  Cewu Lu  Dahua Lin  Jiangmiao Pang 
Shanghai AI Laboratory Shanghai Jiao Tong University The University of Hong Kong
The Chinese University of Hong Kong Tsinghua University

🤖 Demo

📋 Contents

1. About
2. News
3. Getting Started
4. Model and Benchmark
5. TODO List
6. Citation
7. License
8. Acknowledgements

🏠 About

In the realm of computer vision and robotics, embodied agents are expected to explore their environment and carry out human instructions. This necessitates the ability to fully understand 3D scenes given their first-person observations and contextualize them into language for interaction. However, traditional research focuses more on scene-level input and output setups from a global view. To address the gap, we introduce EmbodiedScan, a multi-modal, ego-centric 3D perception dataset and benchmark for holistic 3D scene understanding. It encompasses over 5k scans encapsulating 1M ego-centric RGB-D views, 1M language prompts, 160k 3D-oriented boxes spanning over 760 categories, some of which partially align with LVIS, and dense semantic occupancy with 80 common categories. Building upon this database, we introduce a baseline framework named Embodied Perceptron. It is capable of processing an arbitrary number of multi-modal inputs and demonstrates remarkable 3D perception capabilities, both within the two series of benchmarks we set up, i.e., fundamental 3D perception tasks and language-grounded tasks, and in the wild.

🔥 News

• [2024-03] We first release the data and baselines for the challenge. Please fill in the form to apply for downloading the data and try our baselines. Welcome any feedback!
• [2024-02] We will co-organize Autonomous Grand Challenge in CVPR 2024. Welcome to try the Multi-View 3D Visual Grounding track! We will release more details about the challenge with the baseline after the Chinese New Year.
• [2023-12] We release the paper of EmbodiedScan. Please check the webpage and view our demos!

📚 Getting Started

Installation

We test our codes under the following environment:

• Ubuntu 20.04
• NVIDIA Driver: 525.147.05
• CUDA 12.0
• Python 3.8.18
• PyTorch 1.11.0+cu113
• PyTorch3D 0.7.2

1. Clone this repository.

git clone https://github.com/OpenRobotLab/EmbodiedScan.git
cd EmbodiedScan

2. Create an environment and install PyTorch.

conda create -n embodiedscan python=3.8 -y  # pytorch3d needs python>3.7
conda activate embodiedscan
# Install PyTorch, for example, install PyTorch 1.11.0 for CUDA 11.3
# For more information, please refer to https://pytorch.org/get-started/locally/
conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch

3. Install EmbodiedScan.

# We plan to make EmbodiedScan easier to install by "pip install EmbodiedScan".
# Please stay tuned for the future official release.
# Make sure you are under ./EmbodiedScan/
# This script will install the dependencies and EmbodiedScan package automatically.
# use [python install.py run] to install only the execution dependencies
# use [python install.py visual] to install only the visualization dependencies
python install.py all  # install all the dependencies

Note: The automatic installation script make each step a subprocess and the related messages are only printed when the subprocess is finished or killed. Therefore, it is normal to seemingly hang when installing heavier packages, such as Mink Engine and PyTorch3D.

BTW, from our experience, it is easier to encounter problems when installing these two packages. Feel free to post your questions or suggestions during the installation procedure.

Data Preparation

Please refer to the guide for downloading and organization.

We will update the authorization approach and release remaining data afterward. Please stay tuned.

Tutorial

We provide a simple tutorial here as a guideline for the basic analysis and visualization of our dataset. Welcome to try and post your suggestions!

Demo Inference

We provide a demo for running EmbodiedScan's model on a sample scan. Please download the raw data from Google Drive or BaiduYun and refer to the notebook for more details.

📦 Model and Benchmark

Model Overview

Embodied Perceptron accepts RGB-D sequence with any number of views along with texts as multi-modal input. It uses classical encoders to extract features for each modality and adopts dense and isomorphic sparse fusion with corresponding decoders for different predictions. The 3D features integrated with the text feature can be further used for language-grounded understanding.

Training and Evaluation

We provide configs for different tasks here and you can run the train and test script in the tools folder for training and inference. For example, to train a multi-view 3D detection model with pytorch, just run:

python tools/train.py configs/detection/mv-det3d_8xb4_embodiedscan-3d-284class-9dof.py --work-dir=work_dirs/mv-3ddet --launcher="pytorch"

Or on the cluster with multiple machines, run the script with the slurm launcher following the sample script provided here.

NOTE: To run the multi-view 3D grounding experiments, please first download the 3D detection pretrained model to accelerate its training procedure. After downloading the detection checkpoint, please check the path used in the config, for example, the load_from here, is correct.

To inference and evaluate the model (e.g., the checkpoint work_dirs/mv-3ddet/epoch_12.pth), just run the test script:

python tools/test.py configs/detection/mv-det3d_8xb4_embodiedscan-3d-284class-9dof.py work_dirs/mv-3ddet/epoch_12.pth --launcher="pytorch"

Inference and Submit your Results

We preliminarily support format-only inference for multi-view 3D visual grounding. To achieve format-only inference during test, just set format_only=True in test_evaluator in the corresponding config like here. Then just run the test script like:

python tools/test.py configs/grounding/mv-grounding_8xb12_embodiedscan-vg-9dof.py work_dirs/mv-grounding/epoch_12.pth --launcher="pytorch"

The prediction file will be saved to ./test_results.json in the current directory. You can also set the result_dir in test_evaluator to specify the directory to save the result file.

Finally, to pack the prediction file into the submission format, please modify the script tools/submit_results.py according to your team information and saving paths, and run:

python tools/submit_results.py

Then you can submit the resulting pkl file to the test server (to go live by the end of March) and wait for the lottery :)

We also provide a sample script tools/eval_script.py for evaluating the submission file and you can check it by yourself to ensure your submitted file has the correct format.

Benchmark

We preliminarily provide several baseline results here with their logs and pretrained models.

Note that the performance is a little different from the results provided in the paper because we re-split the training set as the released training and validation set while keeping the original validation set as the test set for the public benchmark.

Multi-View 3D Detection

Method	Input	[email protected]	[email protected]	[email protected]	[email protected]	Download
Baseline	RGB-D	15.22	52.23	8.13	26.66	Model, Log

Multi-View 3D Visual Grounding

Method	[email protected]	[email protected]	Download
Baseline-Mini	33.59	14.40	Model, Log
Baseline-Mini (w/ FCAF box coder)	-	-	-
Baseline-Full	-	-	-

Please see the paper for more details of our two benchmarks, fundamental 3D perception and language-grounded benchmarks. This dataset is still scaling up and the benchmark is being polished and extended. Please stay tuned for our recent updates.

📝 TODO List

• Release the paper and partial codes for datasets.
• Release EmbodiedScan annotation files.
• Release partial codes for models and evaluation.
• Polish dataset APIs and related codes.
• Release Embodied Perceptron pretrained models.
• Release multi-modal datasets and codes.
• Release codes for baselines and benchmarks.
• Full release and further updates.

🔗 Citation

If you find our work helpful, please cite:

@inproceedings{wang2023embodiedscan,
    title={EmbodiedScan: A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI},
    author={Wang, Tai and Mao, Xiaohan and Zhu, Chenming and Xu, Runsen and Lyu, Ruiyuan and Li, Peisen and Chen, Xiao and Zhang, Wenwei and Chen, Kai and Xue, Tianfan and Liu, Xihui and Lu, Cewu and Lin, Dahua and Pang, Jiangmiao},
    year={2024},
    booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
}

If you use our dataset and benchmark, please kindly cite the original datasets involved in our work. BibTex entries are provided below.

Dataset BibTex

@inproceedings{dai2017scannet,
  title={ScanNet: Richly-annotated 3D Reconstructions of Indoor Scenes},
  author={Dai, Angela and Chang, Angel X. and Savva, Manolis and Halber, Maciej and Funkhouser, Thomas and Nie{\ss}ner, Matthias},
  booktitle = {Proceedings IEEE Computer Vision and Pattern Recognition (CVPR)},
  year = {2017}
}

@inproceedings{Wald2019RIO,
  title={RIO: 3D Object Instance Re-Localization in Changing Indoor Environments},
  author={Johanna Wald, Armen Avetisyan, Nassir Navab, Federico Tombari, Matthias Niessner},
  booktitle={Proceedings IEEE International Conference on Computer Vision (ICCV)},
  year = {2019}
}

@article{Matterport3D,
  title={{Matterport3D}: Learning from {RGB-D} Data in Indoor Environments},
  author={Chang, Angel and Dai, Angela and Funkhouser, Thomas and Halber, Maciej and Niessner, Matthias and Savva, Manolis and Song, Shuran and Zeng, Andy and Zhang, Yinda},
  journal={International Conference on 3D Vision (3DV)},
  year={2017}
}

📄 License

This work is under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

👏 Acknowledgements

• OpenMMLab: Our dataset code uses MMEngine and our model is built upon MMDetection3D.
• PyTorch3D: We use some functions supported in PyTorch3D for efficient computations on fundamental 3D data structures.
• ScanNet, 3RScan, Matterport3D: Our dataset uses the raw data from these datasets.
• ReferIt3D: We refer to the SR3D's approach to obtaining the language prompt annotations.
• SUSTechPOINTS: Our annotation tool is developed based on the open-source framework used by SUSTechPOINTS.