We offer two splits for the PartImageNet dataset now.

• PartImageNet_Seg, which is used in the Semantic Part Segmentation and Object Segmentation in the paper (Sec 4.1, 4.2). In this split, train/val/test sets share the same classes but contain different images like standard datasets. It is also used in Compositor: Bottom-up Clustering and Compositing for Robust Part and Object Segmentation.
• PartImageNet_OOD, which is used in the Few-shot Learning experiments in the paper (Sec 4.3). In this split, train/val/test sets have different classes so it is suitable to conduct research on out-of-distribution and few-shot learning problems. (This is also the original released split.)

Our annotations strictly follow the coco style so it should be easy to use the cocoapi for visulization the images and annotations.

If you find our work helpful in your research, please cite it as:

@article{he2021partimagenet,
  title={PartImageNet: A Large, High-Quality Dataset of Parts},
  author={He, Ju and Yang, Shuo and Yang, Shaokang and Kortylewski, Adam and Yuan, Xiaoding and Chen, Jie-Neng and Liu, Shuai and Yang, Cheng and Yuille, Alan},
  journal={arXiv preprint arXiv:2112.00933},
  year={2021}
}

PartImageNet is a large, high-quality dataset with part segmentation annotations. It consists of 158 classes from ImageNet with approximately 24′000 images. The classes are grouped into 11 super-categories and the parts split are designed according to the super-category as shown below. The number in the brackets after the category name indicates the total number of classes of the category.

The statistics of train/val/test split is shown below.

For more detailed statistics, please check out our paper.

PartImageNet has broad potential in and can be benefit to numerious research fields while we simply explore its usage in Part Discovery, Few-shot Learning and Semantic Segmentation in the paper. We hope that with the propose of the PartImageNet, we could attarct more attention to the part-based models and yield more interesting works.