Uni3D: Exploring Unified 3D Representation at Scale

Junsheng Zhou^1,2*, Jinsheng Wang^1*, Baorui Ma^1*, Yu-Shen Liu², Tiejun Huang^1,3, Xinlong Wang¹

¹BAAI, ²THU, ³PKU
^* Equal Contribution

ICLR 2024 (Spotlight)

We present Uni3D, a unified and scalable 3D pretraining framework for large-scale 3D representation learning, and explore its limits at the scale of one billion parameters. Uni3D uses a 2D initialized ViT end-to-end pretrained to align the 3D point cloud features with the image-text aligned features. Via the simple architecture and pretext task, Uni3D can leverage abundant 2D pretrained models as initialization and image-text aligned models as the target, unlocking the great potential of 2D models and scaling-up strategies to the 3D world. We efficiently scale up Uni3D to one billion parameters, and set new records on a broad range of 3D tasks.

Schedule

We are committed to open-sourcing Uni3D related materials, including:

• Extended Uni3D to a 3D metric (Uni3D-score) for enhanced semantic coherence in text-to-3D tasks. For details, see GeoDream.
• The weights of models range from 6M to 1B parameters.
• Evaluation code
• Evaluation data
• Pretraining code
• Pretraining data

We hope to foster the growth of our community through open-sourcing and promoting collaboration👬. Let's step towards multimodal intelligence together🍻.

Installation

Clone this repository and install the required packages:

git clone https://github.com/baaivision/Uni3D.git
cd Uni3D

conda create -n uni3d python=3.8
conda activate uni3d
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

pip install -r requirements.txt

# install pointnet2 extensions from https://github.com/erikwijmans/Pointnet2_PyTorch
pip install "git+git://github.com/erikwijmans/Pointnet2_PyTorch.git#egg=pointnet2_ops&subdirectory=pointnet2_ops_lib"

Core packages:

• Pytorch version 2.0.1
• open-clip-torch version 2.20.0
• timm version 0.9.7
• DeepSpeed version 0.10.3
• Open3D version 0.17.0

Model Zoo

Model	Training Data	Objaverse-LVIS Top1 (Top5)	ModelNet40 Top1 (Top5)	ScanObjectNN Top1 (Top5)
Uni3d-B	Ensembled w/o LVIS	45.9 (74.8)	86.1 (98.7)	61.7 (89.5)
Uni3d-B	Ensembled	51.7 (80.8)	86.3 (97.9)	63.8 (90.2)
Uni3d-L	Ensembled w/o LVIS	46.2 (74.7)	86.6 (97.8)	58.4 (90.1)
Uni3d-L	Ensembled	53.1 (81.5)	86.3 (98.3)	58.2 (89.4)
Uni3d-g	Ensembled w/o LVIS	47.2 (76.1)	86.8 (98.4)	66.5 (90.1)
Uni3d-g	Ensembled	53.5 (82.0)	87.3 (99.2)	63.9 (91.7)
Uni3d-g 🔥	Ensembled	55.3 (82.9)	88.2 (99.3)	65.3 (92.7)

Evaluation of Zero-shot 3D classification

We evaluate the zero-shot 3D classification performance on three datasets: Objaverse-LVIS, ModelNet40 and ScanObjectNN.

1. Please refer to DATASETS.md for evaluation dataset preparation.
2. [Recommended 🤗] Download the clip model and put it in /path/to/clip_model folder.
3. Download model zoo weights and put them in /path/to/checkpoints folder.
4. Run bash scripts/inference.sh [scale] to evaluate the model on the above datasets, e.g., bash scripts/inference.sh giant.

Pre-training

1. Please refer to DATASETS.md for pre-train dataset preparation.
2. [Recommended 🤗] Download the clip model and put it in /path/to/clip_model folder.
3. [Recommended 🤗] Download the initialization model and put it in /path/to/init_model folder.
4. Run bash scripts/pretrain.sh to pre-train the model on ensemble datasets.

Visualization

Open-world Understanding

One-shot Part Segmentation

Point Cloud Painting

Cross-modal Retrieval

Acknowledgement

Uni3D is built using the awesome EVA, OpenCLIP, timm, DeepSpeed, ULIP and OpenShape.

Citation

@inproceedings{zhou2023uni3d,
  title={Uni3d: Exploring unified 3d representation at scale},
  author={Zhou, Junsheng and Wang, Jinsheng and Ma, Baorui and Liu, Yu-Shen and Huang, Tiejun and Wang, Xinlong},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2024}
}