Description
This dataset is developed for benchmarking domain-informed neural networks in the task of underwater moving object detection. The data samples are generated using the Unity Game Engine simulator.
The dataset consists of 1000 independent simulation runs, each creating a video segment. Each video is provided as a sequence of image frames, accompanied by time-synchronized ground-truth data describing the movement of the target object. More specifically, the dataset provides the object’s position, velocity, and acceleration over time, as well as frame timestamps.
YOLO bounding-box annotations are provided per frame and stored in a single JSON file for each video sequence. The images are arranged in folders corresponding to individual video sequences. All images are resized to a resolution of 1920×1080 pixels.
The dataset is intended for research and evaluation of computer vision methods that combine visual information with physics-based knowledge in underwater scenarios.
The dataset consists of 1000 independent simulation runs, each creating a video segment. Each video is provided as a sequence of image frames, accompanied by time-synchronized ground-truth data describing the movement of the target object. More specifically, the dataset provides the object’s position, velocity, and acceleration over time, as well as frame timestamps.
YOLO bounding-box annotations are provided per frame and stored in a single JSON file for each video sequence. The images are arranged in folders corresponding to individual video sequences. All images are resized to a resolution of 1920×1080 pixels.
The dataset is intended for research and evaluation of computer vision methods that combine visual information with physics-based knowledge in underwater scenarios.
| Date made available | 3 Feb 2026 |
|---|---|
| Publisher | TU Delft - 4TU.ResearchData |
Cite this
- DataSetCite