More Text, Less Point: Towards 3D Data-Efficient Point-Language Understanding

Yuan Tang*  Xu Han*  Xianzhi Li^✝  Qiao Yu  Jinfeng Xu  Yixue Hao  Long Hu  Min Chen
Huazhong University of Science and Technology South China University of Technology

AAAI 2025

📋 Contents

• 🔍 Overview
• 📦 Training and Evaluation
• 🔗 Citation
• 📄 License
• 📚 Related Work
• 👏 Acknowledgements

🔍 Overview

• We introduce a new task of 3D data-efficient point-language understanding, aiming to enable LLMs to achieve robust 3D understanding with minimal 3D data.
• We propose GreenPLM to tackle this 3D data-limited task from a novel perspective, enhancing point-LLM alignment with more free-text data.
• we introduce a 6M T3D dataset, design a 3-stage training strategy, and present a 0M-Pooling module for token pooling.
• We introduce the Accuracy-to-3D-Data Ratio (A3DR) to measure the efficiency of 3D data usage and establish an evaluation benchmark based on open-source LLMs.
• GreenPLM outperforms previous models using only 12% of 3D data and even surpasses GPT4Point (660K 3D data) using only text, demonstrating superior 3D data efficiency.

📦 Training-and-Evaluation

Download project

The code, weights, and dataset of the project have already been uploaded to Hugging Face. Simply download them once to get started with the project.

Install Environment

Enter the project directory and execute the following command:

conda create -n greenplm python=3.10 -y
conda activate greenplm
bash  envInstall.sh

Project Directory Introduction

• ./greenplm/release contains the paper's weights, training scripts, and testing scripts.
• ./pretrained_weight stores the pre-trained weights required for the training and testing phases of the project.
• ./lava-vicuna_2024_4_Phi-3-mini-4k-instruct is the weight directory for Phi-3.
• ./dataset/T3D is the 6M T3D dataset proposed in this project.
• ./dataset/T3D/stage_1/brief_1M_caption.json is the dataset from T3D for Stage I.
• ./dataset/T3D/stage_2/stage_2_data_210k.json is the dataset from T3D for Stage II.

Dataset Preparation

./dataset/Objaverse/8192_npy.zip contains the point cloud data from Objaverse that is required for this project. To unzip the dataset:

unzip ./dataset/Objaverse/8192_npy.zip -d ./dataset/Objaverse/

Inference

Paper Weights

GreenPLM-0

The model trained only on text data, i.e., (Stage I & Stage II).

bash ./release/paper/scripts/test/release_stage_2.sh

The output JSON results are saved in ./release/paper/result_json/stage_2.

GreenPLM

The model trained on a small amount of 3D data, i.e., (Stage I & Stage II & Stage III).

bash ./release/paper/scripts/test/release_stage_3.sh

The output JSON results are saved in ./release/paper/result_json/stage_3.

Weights Using All T3D Dataset

bash ./release/5M_data_seting/scripts/test/release_5M_stage_2.sh

bash ./release/5M_data_seting/scripts/test/release_5M_stage_3.sh

Evaluation

Using LLM

• You can get the DASHSCOPE_API_KEY from aliyun. The evaluation may require 9 CNY (~ 1.3 USD).
• If you have enough GPU resources, you can also build your own Qwen2-72B-Instruct service, following the Qwen2. Then evaluate the results for free!

1. Evaluate the open vocabulary classification on objaverse

export PYTHONPATH=$PWD
export DASHSCOPE_API_KEY=sk-xxx
python ./pointllm/eval/evaluator_opensource_llm_QwenAPI.py  \
        --results_path /path/to/evaluation/PointLLM_brief_description_val_200_GT_Objaverse_classification_prompt0.json  \
        --eval_type open-free-form-classification  \
        --model_type qwen2-72b-instruct \
        --parallel --num_workers 4

export PYTHONPATH=$PWD
export DASHSCOPE_API_KEY=sk-xxx
python ./pointllm/eval/evaluator_opensource_llm_QwenAPI.py  \
        --results_path /path/to/evaluation/PointLLM_brief_description_val_200_GT_Objaverse_classification_prompt1.json  \
        --eval_type open-free-form-classification  \
        --model_type qwen2-72b-instruct \
        --parallel --num_workers 4

2. Evaluate the close-set zero-shot classification on ModelNet40

export PYTHONPATH=$PWD
export DASHSCOPE_API_KEY=sk-xxx
python ./pointllm/eval/evaluator_opensource_llm_QwenAPI.py  \
    --results_path /path/to/evaluation/ModelNet_classification_prompt0.json  \
    --eval_type modelnet-close-set-classification  \
    --model_type qwen2-72b-instruct \
    --parallel --num_workers 4

export PYTHONPATH=$PWD
export DASHSCOPE_API_KEY=sk-xxx
python ./pointllm/eval/evaluator_opensource_llm_QwenAPI.py  \
    --results_path /path/to/evaluation/ModelNet_classification_prompt1.json  \
    --eval_type modelnet-close-set-classification  \
    --model_type qwen2-72b-instruct \
    --parallel --num_workers 4

3. Evaluate the object captioning on objaverse

export PYTHONPATH=$PWD
export DASHSCOPE_API_KEY=sk-xxx
python ./pointllm/eval/evaluator_opensource_llm_QwenAPI.py  \
        --results_path /path/to/evaluation/PointLLM_brief_description_val_200_GT_Objaverse_captioning_prompt2.json  \
        --eval_type object-captioning  \
        --model_type qwen2-72b-instruct \
        --parallel --num_workers 4

Traditional Metric Evaluation

For the object captioning task, run the following command to evaluate model outputs with traditional metrics Sentence-BERT and SimCSE.

CUDA_VISIBLE_DEVICES=0 python pointllm/eval/traditional_evaluator.py --results_path /path/to/evaluation/PointLLM_brief_description_val_200_GT_Objaverse_captioning_prompt2.json

Training

Stage I

bash ./release/paper/scripts/train/1.sh

Stage II: GreenPLM-0

bash ./release/paper/scripts/train/2.sh

Stage III: GreenPLM

bash ./release/paper/scripts/train/3.sh

bash ./release/5M_data_seting/scripts/train/2.sh

bash ./release/5M_data_seting/scripts/train/3.sh

Note: You can modify the --output_dir argument in the scripts to set the output directory for the trained weights.

🔗 Citation

If you find our work helpful, please consider citing:

@inproceedings{tang2025more,
  title={More text, less point: Towards 3d data-efficient point-language understanding},
  author={Tang, Yuan and Han, Xu and Li, Xianzhi and Yu, Qiao and Xu, Jinfeng and Hao, Yixue and Hu, Long and Chen, Min},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={39},
  number={7},
  pages={7284--7292},
  year={2025}
}

📄 License

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

📚 Related Work

Together, Let's make LLM for 3D great!

• Point-Bind & Point-LLM: aligns point clouds with Image-Bind to reason multi-modality input without 3D-instruction data training.
• 3D-LLM: employs 2D foundation models to encode multi-view images of 3D point clouds.
• PointLLM: employs 3D point clouds with LLaVA.
• ShapeLLM: combines a powerful point cloud encoder with LLM for embodied scenes.
• MiniGPT-3D : takes the first step toward efficient 3D-LLM, requiring only a single RTX 3090 GPU and one day of training time.

🌟 Star History

👏 Acknowledgements

We would like to thank the authors of PointLLM, Uni3D, Phi-3, and LLaVA-pp for their great works and repos.