This repo hosts the processed data of the ABO dataset for the paper An Object is Worth 64x64 Pixels: Generating 3D Object via Image Diffusion. Please refer to the project homepage, arxiv page and github repo for more details.

Dataset details

We first download the .glb ABO shapes, then we turn the .glb files into 1024x1024x12 object images using Blender 4.0. We set the maximum number of patches to 64 and set the margin to be 2%. The 1024 resolution data is in the data/ folder, and its zipped archive is stored in omages_ABO_p64_m02_1024.tar_partaa and omages_ABO_p64_m02_1024.tar_partab. Please refer to our GitHub repository for instructions on downloading, combining, and extracting the files.

Then, we downsample the 1024 resolution omages to 64 resolution using sparse pooling described in the paper. And we put everything together into the 'df_p64_m02_res64.h5', where the dataset loader will read data items from it.

Citation Information

@misc{yan2024objectworth64x64pixels,
      title={An Object is Worth 64x64 Pixels: Generating 3D Object via Image Diffusion}, 
      author={Xingguang Yan and Han-Hung Lee and Ziyu Wan and Angel X. Chang},
      year={2024},
      eprint={2408.03178},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2408.03178}, 
}
@misc{collins2022abodatasetbenchmarksrealworld,
      title={ABO: Dataset and Benchmarks for Real-World 3D Object Understanding}, 
      author={Jasmine Collins and Shubham Goel and Kenan Deng and Achleshwar Luthra and Leon Xu and Erhan Gundogdu and Xi Zhang and Tomas F. Yago Vicente and Thomas Dideriksen and Himanshu Arora and Matthieu Guillaumin and Jitendra Malik},