3dlg-hcvc/DuoduoCLIP-data

Dataset Card for DuoduoCLIP

In this data repo we provide the data used in the paper Duoduo CLIP: Efficient 3D Understanding with Multi-View Images. The data usage and code can be found in the github repo.

Note: We provide the lvis evaluation data in the initial release, we will soon upload the data and process scripts required for training.

Dataset Details

Dataset Sources

Multi-view images of 3D objects were used in training our models. A majority of the Objaverse renderings come from the Zero123 paper (MIT license) listed below. The initial release with LVIS images in 3dlg-hcvc/DuoduoCLIP-data/lvis_split are also mostly from Zero123 with some preprocessing. Note we provide this for ease of evaluation. For the other releases we will provide the dataset preprocessing scripts instead and point users towards their download link. For objects not rendered by Zero123 and the other 3 datasets (ABO, ShapeNet and 3D-FUTURE) we render using Zero123's blender script. We thank the authors for providing their dataset and code!

1. Zero123

• Repository: https://github.com/cvlab-columbia/zero123
• Paper: https://arxiv.org/abs/2303.11328

2. Objaverse

• Repository: https://github.com/allenai/objaverse-xl
• Paper: https://arxiv.org/abs/2212.08051

3. ABO

• Repository: https://github.com/jazcollins/amazon-berkeley-objects
• Paper: https://arxiv.org/abs/2110.06199

4. ShapeNet

• Repository: https://huggingface.co/ShapeNet
• Paper: https://arxiv.org/abs/1512.03012

5. 3D-FUTURE

• Website: https://tianchi.aliyun.com/specials/promotion/alibaba-3d-future
• Paper: https://arxiv.org/abs/2009.09633

Embeddings from our Models

We will also provide objaverse embeddings produced by our released models in this repo. In the initital release we provide the objaverse embeddings produced by our Four_1to6F_bs1600_LT6 model. Please see the model card for more details and github repo for usage.