README.md

metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: segment_mask
      dtype: image
    - name: detections
      dtype: string
    - name: referring_expression
      dtype: string
  splits:
    - name: train
      num_bytes: 17651696
      num_examples: 100
  download_size: 17392053
  dataset_size: 17651696
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Paper Link

SOS: Synthetic Object Segments Improve Detection, Segmentation, and Grounding

Code repo

Code for Generation

Citation

@misc{huang2025sossyntheticobjectsegments,
  title={SOS: Synthetic Object Segments Improve Detection, Segmentation, and Grounding},
  author={Weikai Huang and Jieyu Zhang and Taoyang Jia and Chenhao Zheng and Ziqi Gao and Jae Sung Park and Ranjay Krishna},
  year={2025},
  eprint={2510.09110},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2510.09110}
}