ZixunWang/rankseg-benchmark-models
Image Segmentation • Updated
Datasets:
Access the RankSEG benchmark probabilities
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset contains model probabilities and ground-truth labels derived from third-party datasets (PASCAL VOC 2012, Cityscapes, ADE20K, DeepGlobe Land Cover, KiTS19 and LiTS). It is provided for non-commercial research only. By requesting access, you agree to use it only for non-commercial research and to comply with the terms of each original dataset, including any restrictions on redistribution.
Log in or Sign Up to review the conditions and access this dataset content.
RankSEG Benchmark Probabilities
Precomputed test-set probabilities and ground-truth labels used by the Quick suite of
rankseg-benchmark. They let you
compare RankSEG decoders with argmax without
running a segmentation model. The models that produced them are in
ZixunWang/rankseg-benchmark-models.
Access
The dataset is gated. Log in to Hugging Face, accept the conditions on this page, and
authenticate locally (huggingface-cli login or the HF_TOKEN environment variable)
before running rankseg-bench quick.
Contents
| Directory | Source data | Inputs | Classes |
|---|---|---|---|
pascal_voc/ |
PASCAL VOC 2012 val | 1,449 images | 21 |
cityscapes/ |
Cityscapes val | 500 images | 19 |
ade20k/ |
ADE20K val | 2,000 images | 150 |
deepglobe_land/ |
DeepGlobe Land Cover, test split | 163 images | 6 (+ unknown, index 6) |
kits/fold{0-4}/ |
KiTS19, 5-fold cross-validation | 2D CT slices | 2 |
lits/fold{0-4}/ |
LiTS, 5-fold cross-validation | 2D CT slices | 2 |
Each row has these fields:
probs: NumPy.npybytes, float16 softmax probabilities of shape(classes, H, W);label: NumPy.npybytes, ground-truth class indices of shape(H, W); ignored pixels use 255 (6 for DeepGlobe Land);image_file,dataset, and for KiTS and LiTScase_idandslice_index.
License and terms
This dataset is released under CC BY-NC-SA 4.0 for non-commercial research use only, and is provided as is, without warranty. It is not intended for clinical use.
The probabilities and labels are derived from the datasets below. You must also comply with each original dataset's terms, some of which are stricter than CC BY-NC-SA 4.0; for example, Cityscapes may not be used commercially or redistributed.
| Subset | Original dataset and terms |
|---|---|
pascal_voc/ |
PASCAL VOC 2012 |
cityscapes/ |
Cityscapes license |
ade20k/ |
ADE20K terms of use |
deepglobe_land/ |
DeepGlobe 2018 challenge |
kits/ |
KiTS19 (CC BY-NC-SA 4.0) |
lits/ |
LiTS challenge (CC BY-NC-SA 4.0) |
Citation
@inproceedings{wang2025rankseg,
title={RankSEG-RMA: An Efficient Segmentation Algorithm via Reciprocal Moment Approximation},
author={Wang, Zixun and Dai, Ben},
booktitle={Advances in Neural Information Processing Systems},
year={2025}
}
@inproceedings{wang2026rankseg,
title={On the Relaxation of Conditional Independence Assumption for Image Segmentation},
author={Wang, Zixun and Dai, Ben},
booktitle={Advances in Neural Information Processing Systems},
year={2026}
}
- Downloads last month
- 266
Models trained or fine-tuned on ZixunWang/rankseg-benchmark