TUBcertainty multi-annotator study - dataset

Official data & annotations for the ECCVW paper A Multi-Annotator Study of Segmentation Noise and Uncertainty in Turbid Underwater Images (https://arxiv.org/abs/2608.15363).

The dataset includes annotations of 15 underwater images, with each image annotated by multiple annotators. The images and golden GT are taken from the TUB dataset https://huggingface.co/datasets/vapaau/TUB. The annotations were collected in a user study, where each annotator was asked to label all objects in 3 images (cf. the paper for details)

Each study participant is assigned an anonymous user_id, which we keep consistent across the images they annotated. We provide the user_id and other user metadata as part of the dataset (no personally identifying information).

Here we provide the images, instance segmentation annotations (polygons), binary segmentation masks, and user metadata. + some helper scripts to visualize the annotations, or compare them to the GT.

Folder structure

TUBcertainty/
├── images/                          # The 15 RGB images (1920x1080 .jpg)
├── instance_segmentation/
│   ├── ground_truth_15imgs.json     # GT, original format
│   ├── ground_truth_15imgs_coco.json  # GT, COCO format
│   ├── multiannotator_15imgs.json   # All annotators, one entry per image
│   └── multiannotator_15imgs_coco/  # All annotators, one COCO file per annotator
│       └── <user_id>.json
├── binary_segmentation/             # Foreground/background masks, for semantic segmentation (.png)
│   ├── <image_name>_GT.png
│   └── <image_name>_<user_id>.png
├── user_metadata_per_image.csv      # Per annotator and image: annotation time, image number, display settings, questionnaire answers, ...
├── visualize_gt_vs_annotator.py
├── visualize_per_image.py
└── visualize_per_user.py

Data

images/

The 15 images that were annotated in the user study, named Exp<N>_Cam<N>_Stage<NN>.jpg (5 scenes, 3 turbidity levels per scene). Note that these are directly taken from the TUB dataset https://huggingface.co/datasets/vapaau/TUB

instance_segmentation/

All coordinates are in image pixels (origin at the top-left corner).

• ground_truth_15imgs.json: {image_name: [{"bbox": [x, y, w, h], "segmentation": [[x1, y1, x2, y2, ...]]}, ...]}, one entry per GT object.
• ground_truth_15imgs_coco.json: the same GT as a standard COCO instance segmentation file. There is a single category (object), area is the mask area in pixels, and image ids are shared with the per-annotator COCO files.
• multiannotator_15imgs.json: {image_name: [annotator, ...]}. Each annotator entry has user_id, image_name, image_num, user_group, turbidity_level and a list of polygons. Each polygon has polygon_id, polygon_pts (list of {"x", "y"}), polygon_center_x, polygon_center_y and polygon_confidence (the annotator's own confidence). image_num is either 1,2 or 3 (if this is the first, second or third image in the annotator sequence) and user_group is the group to which the annotator was assigned (cf. paper).
• multiannotator_15imgs_coco/<user_id>.json: the annotations of one annotator as a COCO file, containing only the images that annotator labelled. Image ids match the GT COCO file, so any of these files can be evaluated directly against ground_truth_15imgs_coco.json with pycocotools. Each annotation additionally has polygon_id and confidence fields, which COCO tools ignore.

binary_segmentation/

One boolean mask per image for the GT (<image_name>_GT.png) and per annotator (<image_name>_<user_id>.png), at the image resolution.

user_metadata_per_image.csv

Metadata for each annotator and image.

Scripts

Script	Purpose
visualize_gt_vs_annotator.py	Sanity-check/demo for one image and one annotator. Plots the GT, plots the annotator's polygons, computes COCO metrics (segm and bbox AP/AR) of the annotator against the GT, then computes the binary IoU between the two masks and plots both masks in one overlay (green: both, red: GT only, blue: annotator only).
visualize_per_image.py	Plots all annotators' polygons for one image, one subplot per annotator. Usage: python visualize_per_image.py [image_name] (random image if omitted).
visualize_per_user.py	Plots all annotations of one annotator across the images they labelled, one subplot per image. Usage: python visualize_per_user.py [user_id] (random user if omitted).

The visualization scripts save their figure as example_*.png in the current directory and also show it.

Citation

If you use this dataset in your research, please cite our study, as well as the original TUB dataset:

@inproceedings{tubcertainty-multiannotator-study,
title = "A Multi-Annotator Study of Segmentation Noise and Uncertainty in Turbid Underwater Images",
author = "Galadrielle Humblot-Renaux and Vasiliki Ismiroglou and Malte Pedersen",
year = "2026",
month = aug,
language = "English",
booktitle = "Computer Vision -- ECCV 2026 Workshops",
publisher = "Springer",
}

@InProceedings{tub-dataset,
author="Ismiroglou, Vasiliki and Bengtson, Stefan H. and Benos, Tasos and Moeslund, Thomas B. and Pedersen, Malte",
title="Beyond Aesthetics: Quantifying Information Loss in Turbid Scenes",
booktitle="Computer Vision -- ECCV 2026",
year="2026",
publisher="Springer Nature Switzerland",
address="Cham",
pages="492--509",
isbn="978-3-032-36984-0"
}