|
|
--- |
|
|
task_categories: |
|
|
- image-segmentation |
|
|
tags: |
|
|
- AI |
|
|
- SAM |
|
|
size_categories: |
|
|
- 10K<n<100K |
|
|
--- |
|
|
# UncertSAM Benchmark |
|
|
|
|
|
## Dataset Summary |
|
|
|
|
|
**UncertSAM** is a curated multi-domain benchmark designed to stress-test segmentation foundation models (specifically the Segment Anything Model, SAM) under challenging conditions. It comprises eight datasets spanning diverse visual degradations and environments, including shadows, transparency, camouflage, and medical imaging. |
|
|
|
|
|
This benchmark was introduced in the paper **"Towards Integrating Uncertainty for Domain-Agnostic Segmentation"** to investigate whether uncertainty quantification can enhance model generalisability and robustness in shifted or limited-knowledge domains. |
|
|
|
|
|
--- |
|
|
|
|
|
## Included Datasets and Licensing |
|
|
|
|
|
**⚠️ Important:** The UncertSAM benchmark aggregates existing datasets. While the benchmark curation allows for domain-agnostic analysis, users must adhere to the original licenses of the individual subsets. Below is the licensing information for each component as detailed in the paper: |
|
|
|
|
|
| Dataset | Source / Reference | License | Domain / Challenge | |
|
|
| :--- | :--- | :--- | :--- | |
|
|
| **BIG** | Cheng et al. [2020] | **Research Only** | Fine-grained salient objects | |
|
|
| **COIFT** | Liew et al. [2021] | **Attribution-NonCommercial 4.0** | Fine-grained salient objects | |
|
|
| **COD10K-v3** | Fan et al. [2022] | **Research Only** | Camouflaged objects | |
|
|
| **MSD Spleen** | Antonelli et al. [2022] | **Attribution-ShareAlike 4.0** | Medical CT scans | |
|
|
| **ISTD** | Wang et al. [2018] | **Research Only** | Shadows | |
|
|
| **SBU** | Vicente et al. [2016] | **Unknown** | Shadows | |
|
|
| **Flare7K** | Dai et al. [2022] | **S-Lab License 1.0** | Lighting artifacts (Flares) | |
|
|
| **Trans10K** | Xie et al. [2021] | **Research Only** | Transparent objects | |
|
|
|
|
|
*Note: For the SBU dataset where the license is listed as "Unknown," please exercise caution and refer to the original publication for terms of use.* |
|
|
|
|
|
--- |
|
|
|
|
|
## Dataset Statistics |
|
|
|
|
|
*I = number of images, M = number of corresponding segmentation masks per split.* |
|
|
|
|
|
| Dataset | Train (I) | Train (M) | Val (I) | Val (M) | Test (I) | Test (M) | |
|
|
| ------- | --------- | --------- | ------- | ------- | -------- | -------- | |
|
|
| Trans | 7,289 | 16,443 | 1,560 | 3,604 | 1,560 | 3,609 | |
|
|
| MSD | 668 | 668 | 135 | 135 | 135 | 135 | |
|
|
| ISTD | 1,309 | 1,408 | 275 | 295 | 285 | 300 | |
|
|
| COD | 3,505 | 4,049 | 760 | 885 | 750 | 846 | |
|
|
| BIG | 97 | 120 | 19 | 21 | 30 | 36 | |
|
|
| COIFT | 209 | 209 | 41 | 41 | 30 | 30 | |
|
|
| Flare | 140 | 163 | 30 | 33 | 30 | 38 | |
|
|
| SBU | 3,278 | 7,505 | 705 | 1,643 | 705 | 1,592 | |
|
|
|
|
|
--- |
|
|
|
|
|
|
|
|
## Dataset Preprocessing |
|
|
|
|
|
To standardise the benchmarks for evaluation with SAM, several preprocessing steps were applied to the original data. The specific processing methods used for different subsets are described below: |
|
|
|
|
|
### 1\. Connected Component Analysis (CCA) |
|
|
|
|
|
Applied to datasets containing images with multiple disconnected surfaces (potentially valid masks under SAM's entity-part strategy) to separate them into distinct masks. |
|
|
|
|
|
* **Morphological Closing:** Applied twice using a 3x3 kernel to connect small disconnected parts that belong to a larger coherent target. |
|
|
* **Component Extraction:** Connected components are extracted from the closed mask. |
|
|
* **Refinement:** A binary AND operation is performed between the resulting mask and the initial mask to remove artifacts introduced by closing. Components smaller than 1,000 pixels are discarded. |
|
|
|
|
|
### 2\. Colour Coded (CC) Processing |
|
|
|
|
|
For datasets where masks are colour-coded (multiple classes/objects in one mask file), unique RGB values are extracted, and the mask is split into multiple binary targets accordingly. |
|
|
|
|
|
### 3\. CT Scan Processing (MSD Spleen) |
|
|
|
|
|
The original 3D CT volumes (`.nii.gz` format) were processed into 2D images: |
|
|
|
|
|
* **Slicing:** Volumes were sliced along the axial plane, retaining only slices containing foreground labels. |
|
|
* **Normalisation:** Z-score normalisation was applied. |
|
|
* **Clipping:** Voxel intensities were clipped to the [0.05, 99.95] percentile range. |
|
|
|
|
|
## Citation |
|
|
|
|
|
If you use this benchmark in your research, please cite the original paper: |
|
|
|
|
|
```bibtex |
|
|
@article{brouwers2025towards, |
|
|
title={Towards Integrating Uncertainty for Domain-Agnostic Segmentation}, |
|
|
author={Brouwers, Jesse and Xing, Xiaoyan and Timans, Alexander}, |
|
|
journal={arXiv preprint arXiv:2512.23427}, |
|
|
year={2025} |
|
|
} |
|
|
``` |
|
|
|
|
|
Please also ensure you cite the original authors of the individual datasets included in this benchmark. |
