Dataset Card for ImageNet-S

This dataset is an unofficial replica of the ImageNet-S validation set and is uploaded here for ease of use by the AI community. ImageNet-S is a large-scale dataset for semantic segmentation derived from ImageNet. This dataset contains only the validation set with 12,419 images across 919 semantic categories with pixel-level annotations.

Dataset Details

Dataset Description

ImageNet-S is a semantic segmentation dataset built upon the ImageNet-1K classification dataset, designed for large-scale unsupervised and semi-supervised semantic segmentation (LUSS). The dataset provides pixel-level semantic annotations for images across 919 object categories.

The dataset addresses two major challenges in unsupervised segmentation: (i) the lack of a large-scale benchmark for assessing algorithms, and (ii) the difficulty of unsupervised shape representation learning. It features high data diversity and clear task objectives, making it suitable for evaluating un/weakly/fully supervised segmentation methods.

This validation set contains 12,419 images with high-quality semantic segmentation annotations across 919 categories.

• Curated by: LUSSeg Team (Shanghua Gao, Zhong-Yu Li, Ming-Hsuan Yang, Ming-Ming Cheng, Junwei Han, Philip Torr)
• Language(s) (NLP): Not applicable (Computer Vision dataset)
• License: Follows ImageNet license terms

Dataset Sources

• Repository: https://github.com/LUSSeg/ImageNet-S
• Project Page: https://LUSSeg.github.io/
• Paper: Large-scale Unsupervised Semantic Segmentation (TPAMI 2022)

Uses

Direct Use

This validation set is intended for:

• Evaluating semantic segmentation models
• Benchmarking segmentation algorithms across 919 object categories
• Assessing unsupervised and semi-supervised semantic segmentation methods
• Validating transfer learning approaches for segmentation tasks
• Testing unsupervised shape representation learning

Out-of-Scope Use

• Training models (this is a validation set only)
• Instance segmentation tasks (dataset provides semantic, not instance-level masks)
• Panoptic segmentation without additional annotations
• Tasks requiring fine-grained boundary precision beyond the annotation quality
• Commercial applications without proper licensing compliance with ImageNet terms

Dataset Structure

This dataset contains the ImageNet-S validation set:

• Images: 12,419 RGB JPEG images
• Categories: 919 semantic categories
• Annotations: Pixel-level segmentation masks for all images

Data Format

Images: RGB JPEG images of varying resolutions from ImageNet-1K validation set.

Segmentation Masks: PNG format with RGB channels where the ImageNet-S category-ID is encoded as: category_id = R + G * 256

• Ignored regions: labeled as 1000
• "Other" category: labeled as 0

[* IMPORTANT *] Label Column: The "label" field in this replicated version contains the original ImageNet-1K classification label (ranging from 0-999), which corresponds to the image-level category from the original ImageNet dataset. This is distinct from the ImageNet-S segmentation category-ID found in the mask annotations.

Category Mapping

ImageNet-S uses the same tag ID system as ImageNet (e.g., n01440764, n02102040) but includes only 919 of the original 1,000 ImageNet categories. The ImageNet-S category-IDs are obtained by sorting these ImageNet tag IDs alphabetically.

The mapping table between ImageNet-S categories and ImageNet tag IDs is available at:

• ImageNet-S-919 categories

Note: Some ImageNet categories are merged in ImageNet-S-919:

n04356056 → n04355933
n04493381 → n02808440
n03642806 → n03832673
n04008634 → n03773504
n03887697 → n15075141

Dataset Creation

Curation Rationale

ImageNet-S was created to enable large-scale unsupervised semantic segmentation research. While ImageNet provided an enormous resource for classification tasks, segmentation research was limited by smaller datasets like PASCAL VOC and COCO. The creators aimed to:

1. Provide a large-scale benchmark for unsupervised/semi-supervised semantic segmentation
2. Enable research on unsupervised shape representation learning at scale
3. Offer a benchmark with significantly more categories and data diversity than existing segmentation datasets
4. Support multiple evaluation protocols (fully unsupervised, semi-supervised, distance matching)
5. Establish clear baselines and track research progress in LUSS

Source Data

Data Collection and Processing

The source images come from the ImageNet-1K validation set, which contains images collected from the web and organized according to the WordNet hierarchy. For ImageNet-S, high-quality semantic segmentation masks were created through careful annotation processes.

Some ImageNet categories were merged to create 919 meaningful semantic categories for segmentation.

Who are the source data producers?

The original images were collected by the ImageNet team (Princeton University and Stanford University). The semantic segmentation annotations were created by the LUSSeg research team.

Annotations

Annotation process

All 12,419 validation images have precise pixel-level semantic segmentation masks. Annotations are stored in PNG format with RGB channels where the category-ID is encoded as R + G * 256. Ignored regions are labeled as 1000, and the "other" category is labeled as 0.

Who are the annotators?

The annotations were created by the LUSSeg research team through their annotation pipeline.

Personal and Sensitive Information

As the images are sourced from ImageNet, they may contain people, faces, and scenes that could include identifiable individuals or sensitive content. ImageNet has documented known biases and privacy concerns. Users should be aware that:

• Images may contain identifiable people
• No specific anonymization was performed beyond ImageNet's original curation
• Images reflect biases present in web-scraped data from the ImageNet era

Bias, Risks, and Limitations

Biases:

• Inherits geographic, cultural, and demographic biases from ImageNet
• Object categories reflect Western/English-centric taxonomy (WordNet)
• Potential underrepresentation of certain demographics and regions

Technical Limitations:

• Validation set only (no training data included)
• Annotation quality may vary across images
• Some categories have merged classes which may affect fine-grained distinctions
• Segmentation masks may not capture extremely fine-grained details

Risks:

• Privacy concerns due to potentially identifiable individuals
• Perpetuation of biases if used without consideration
• May not generalize well to underrepresented domains

Recommendations

Users should:

• Use this dataset for evaluation purposes only (not for training)
• Be aware of inherited ImageNet biases and limitations
• Consider privacy implications when using images containing people
• Evaluate model performance across diverse demographics and contexts
• Review and comply with ImageNet's terms of use
• Consider fairness and bias mitigation strategies in downstream applications
• Understand the distinction between the "label" column (ImageNet-1K classification) and the segmentation mask category-IDs (ImageNet-S categories)

Citation

BibTeX:

@article{gao2022luss,
  title={Large-scale Unsupervised Semantic Segmentation},
  author={Gao, Shanghua and Li, Zhong-Yu and Yang, Ming-Hsuan and Cheng, Ming-Ming and Han, Junwei and Torr, Philip},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2022}
}

APA:

Gao, S., Li, Z. Y., Yang, M. H., Cheng, M. M., Han, J., & Torr, P. (2022). Large-scale Unsupervised Semantic Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence.

Glossary

• LUSS: Large-scale Unsupervised Semantic Segmentation - the task of segmenting images at scale without labeled training data
• Semantic Segmentation: The task of assigning a category label to each pixel in an image
• ImageNet-1K: The 1,000-class subset of ImageNet used for the ILSVRC classification challenge
• Validation Set: A subset of data used to evaluate model performance
• ImageNet Tag ID: The WordNet synset identifier used in ImageNet (e.g., n01440764)
• Category-ID: The numeric identifier for semantic segmentation categories in ImageNet-S, encoded in the segmentation masks

More Information

For more information, please refer to:

• Original repository: https://github.com/LUSSeg/ImageNet-S
• Project page: https://LUSSeg.github.io/
• ImageNet project: https://www.image-net.org/
• Baseline methods:
  • PASS (Unsupervised): https://github.com/LUSSeg/PASS
  • Semi-supervised: https://github.com/LUSSeg/ImageNetSegModel
  • MMSegmentation config: https://github.com/open-mmlab/mmsegmentation/tree/master/configs/imagenets

Dataset Card Contact

For all questions about ImageNet-S, please contact shgao@live.com or refer to the original GitHub repository: https://github.com/LUSSeg/ImageNet-S