ScanReason

<br> <p align="center"> <h1 align="center"><strong>Empowering 3D Visual Grounding with Reasoning Capabilities</strong></h1> <p align="center"> <strong>ECCV 2024</strong> <br> <a href='https://zcmax.github.io//' target='_blank'>Chenming Zhu</a>&emsp; <a href='https://tai-wang.github.io/' target='_blank'>Tai Wang</a>&emsp; <a href='https://zhangwenwei.cn/' target='_blank'>Wenwei Zhang</a>&emsp; <a href='https://chenkai.site/' target='_blank'>Kai Chen</a>&emsp; <a href='https://xh-liu.github.io//' target='_blank'>Xihui Liu*</a>&emsp; <br> The University of Hong Kong&emsp;Shanghai AI Laboratory <br> </p> </p> <div id="top" align="center">

</div> <!-- ## 🏠 Background <!--  --> <!-- <div style="text-align: center;"> <img src="assets/Fig_Teaser.png" alt="Dialogue_Teaser" width=100% > </div> For an embodied agent, they not only need to be able to understand the 3D environment and complex human instructions, but also localize the target objects for interaction and navigation. Although GPT-4 (GPT-4V) have strong text (multi-modal) reasoning abilities, they lack the ability to directly perceive the 3D scene, understand the 3D spatial relationships and output corresponding target object locations. In this paper, we propose a new task called 3D reasoning grounding and introduce a new benchmark **ScanReason** which provides over 10K question-answer-location pairs from five reasoning types that require the synerization of reasoning and grounding. We further design our approach, ReGround3D, composed of the visual-centric reasoning module empowered by Multi-modal Large Language Model (MLLM) and the 3D grounding module to obtain accurate object locations by looking back to the enhanced geometry and fine-grained details from the 3D scenes. A chain-of-grounding mechanism is proposed to further boost the performance with interleaved reasoning and grounding steps during inference. -->

📦 Benchmark and Model

Benchmark Overview

<p align="center"> <img src="assets/scanreason_benchmark_v2.png" align="center" width="100%"> </p> ScanReason is the first comprehensive and hierarchical 3D reasoning grounding benchmark. We define 5 types of questions depending on which type of reasoning is required: Spatial reasoning and function reasoning require fundamental understanding of the 3D physical world, focusing on objects themselves and inter-object spatial relationships in a 3D scene respectively, and logistic reasoning, emotional reasoning, and safety reasoning are high-level reasoning skills built upon the two fundamental reasoning abilities to address user-centric real-world applications. <!-- <p align="center"> <img src="assets/3d_reasoning_grounding.png" align="center" width="100%"> </p> -->

Model Overview

<p align="center"> <img src="assets/Fig_Method.png" align="center" width="100%"> </p>

🔥 News

• [2023-10-10] We release our pre-version of ScanReason validation benchmark. Download here. The corresponding 3D bounding boxes annotations could be obtained through the object ids from EmbodiedScan.
• [2023-10-01] We release the training and inference codes of ReGround3D.
• [2023-07-02] We release the paper of ScanReason.

Getting Started

1. Installation

• We utilize at least 4 A100 GPU for training and inference.

• We test the code under the following environment:

  • CUDA 11.8
  • Python 3.9
  • PyTorch 2.1.0

• Git clone our repository and creating conda environment:

  git clone https://github.com/ZCMax/ScanReason.git
  conda create -n scanreason python=3.9
  conda activate scanreason
  pip install -r requirements.txt
  

• Follow EmbodiedScan Installation Doc to install embodiedscan series.

• Compile Pointnet2

  cd pointnet2
  python setup.py install --user
  

2. Data Preparation

1. Follow EmbodiedScan Data Preparation Doc to download the raw scan (RGB-D) datasets and modify the VIDEO_FOLDER in train_ds.sh to the raw data path.

2. Download the text annotations from Google Drive and modify the JSON_FOLDER in train_ds.sh to the annotations path, and modify the INFO_FILE data path which is included in the annotations.

3. Training ReGround3D

We provide the slurm training script with 4 A100 GPUs:

./scripts/train_ds.sh

4. Evaluation ReGround3D

After training, you can run the

./scripts/convert_zero_to_fp32.sh 

to convert the weights to pytorch_model.bin file, and then use

./scripts/merge_lora_weights.sh

to merge lora weight and obtain the final checkpoints under ReGround3D-7B.

Finally, run

./scripts/eval_ds.sh

to obtain the grounding results.

📝 TODO List

• [x] First Release.
• [x] Release ReGround3D code.
• [ ] Release ScanReason datasets and benchmark.

📄 License

<a rel="license" href="http://creativecommons.org/licenses/by-nc-sa/4.0/"><img alt="Creative Commons License" style="border-width:0" src="https://i.creativecommons.org/l/by-nc-sa/4.0/80x15.png" /></a> <br /> This work is under the <a rel="license" href="http://creativecommons.org/licenses/by-nc-sa/4.0/">Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License</a>.

👏 Acknowledgements

This repo benefits from LISA, EmbodiedScan, 3D-LLM, LLaVA.