You must agree to the Data License Agreement for SeaClips to access it.
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This data license agreement ("Agreement") between SEA.AI GmbH ("SEA.AI") and you, whether an individual or entity, ("you") governs the use of the SeaClips dataset, provided by SEA.AI ("Dataset"). The Dataset consists color images forming video sequences ("Images"), and annotation files, which describe video-level and frame-level properties and metadata ("Annotations").
You agree to use the Dataset only for purposes expressly permitted by this Agreement and in accordance with any applicable law or regulation in the relevant jurisdictions.
- Purpose limitation. The Dataset serves exclusively for scientific research on maritime obstacle detection in the context of improving those computer vision solutions.
- Scope of license. The dataset may only be used for scientific research on maritime obstacle detection. It may not be used for any other purpose, including commercial purposes as, for example, licensing or selling the data, or using the data with a purpose to procure a commercial gain. SEA.AI gives you the non-exclusive, revocable, and non-transferable right to use the dataset. If a new version of the Dataset is published and upon written request by SEA.AI, you will use the updated version of the Dataset and delete any prior versions.
- Restriction of use. It is prohibited to combine the Dataset with other data sources that allow a (re-)identification or violation of IP-rights. It is prohibited to use the Dataset to identify or attempt to identify any natural person or for de-anonymization.
- Sharing with third parties. Direct sharing of parts or the full Dataset with third parties is prohibited. You may share the results of your algorithm derived from the Dataset in scientific publications as long as the original Dataset cannot be derived from it.
- Warranty and liability. The dataset is provided "AS IS" and with all its faults. SEA.AI is liable only for intent or extreme gross negligence. SEA.AI is not liable for violation of IP or data protection rights.
- Termination. SEA.AI can terminate this contract immediately in case of failure to comply with the terms in this Agreement. Upon termination, you must stop using and destroy all copies of the Dataset.
- Feedback. Any mistakes within the dataset or any re-identification findings can and must be reported immediately to [email protected]. Please use the tag [SeaClips] in the email subject.
- Governing law and jurisdication. Austrian law with jurisdication in Linz or Vienna.
Log in or Sign Up to review the conditions and access this dataset content.
SeaClips: A Video Dataset for Maritime Object Detection
SeaClips is a video-based maritime obstacle detection dataset. 🚢⛵🌊🎬
Dataset Details
General statistics: The dataset consists of 74 videos, with 31k frames and 129k bounding boxes. The videos are stored at 30 FPS and have an average duration of approximately 14 seconds.
Object Categories: There are six vessel and non-vessel categories available, as well as one generic object class (see Table below).
| ID | Name | Description |
|---|---|---|
| 1 | ANIMAL | Birds |
| 2 | BOAT | Motorized marine vehicle |
| 3 | MARINE_MARKER | Signs with fixed position for navigation and information |
| 4 | LEISURE_VEHICLE | Non-motorized vehicle |
| 5 | OBJECT | Visibly present, unrecognizable or not belonging to the other classes |
| 6 | SAILING_VESSEL | Marine vehicle whose primary propulsion is sails |
| 7 | SHIP | Large motorized marine vehicle |
Dataset Owner(s)
SEA.AI GmbH
Dataset Creation Date
12/03/2025
Intended Use
The usage of this dataset is restricted to non-commercial research on maritime obstacle detection to improve safety at sea.
Dataset Structure
The dataset structure follows closely the COCO VID format, as it is used in MMTracking [1].
Folder Structure
The file structure looks like this:
SeaClips/
├─ train/
│ ├─ video_01/
│ │ ├─ 000001.png
│ │ ├─ 000002.png
│ │ ├─ ...
│ ├─ video_02/
│ │ ├─ 000001.png
│ │ ├─ 000002.png
│ │ ├─ ...
│ ├─ .../
├─ val/
│ ├─ video_01/
│ │ ├─ 000001.png
│ │ ├─ 000002.png
│ │ ├─ ...
│ ├─ video_02/
│ │ ├─ 000001.png
│ │ ├─ 000002.png
│ │ ├─ ...
│ ├─ .../
├─ test/
│ ├─ video_01/
│ │ ├─ 000001.png
│ │ ├─ 000002.png
│ │ ├─ ...
│ ├─ video_02/
│ │ ├─ 000001.png
│ │ ├─ 000002.png
│ │ ├─ ...
│ ├─ .../
├─ seaclips-train.json
├─ seaclips-val.json
├─ seaclips-test.json
Annotation Format
The annotation json-files contain dictionaries with the following keys:
| Key | Type | Description |
|---|---|---|
info |
Dict |
General information about the dataset (e.g., version, description, contributor). |
categories |
List[Dict] |
List of available categories. Each category dictionary contains: • id (int): Unique category identifier.• name (str): Category name.• supercategory (str): Higher-level grouping for the category (always None, included for completeness). |
videos |
List[Dict] |
List of videos in the dataset. Each entry includes: • id (int): Unique video identifier.• name (str): Video file name (e.g., "video01").• weather (str): sunny/overcast/cloudy/rainy. • water_reflections (str): yes/no. • sea_state (str): smooth/wavy. • camera_sensor (str): e-CAM/Axis/FLIR. |
images |
List[Dict] |
List of image metadata. Each image dictionary includes: • id (int): Image identifier.• video_id (int): Video identifier to which the frame belongs.• frame_id (int): Frame identifier within the video.• file_name (str): Image filename.• height (int): Image height in pixels.• width (int): Image width in pixels. • weather (str): sunny/overcast/cloudy/rainy. • water_reflections (str): yes/no. • sea_state (str): smooth/wavy. • camera_sensor (str): e-CAM/Axis/FLIR. |
annotations |
List[Dict] |
List of object annotations. Each dictionary includes: • id (int): Annotation ID.• image_id (int): ID of the associated image.• category_id (int): ID of the associated category.• bbox (List[float]): Bounding box [x, y, width, height].• area (float): Area of the object.• iscrowd (int): 0 or 1 indicating crowd annotation (always 0, included for completeness). |
Example Usage and Code
Some example code for data loading and visualization is available here. It can be installed as package to easily integrate into your project.
Bias, Risks, and Limitations
There are three main biases:
- Weather: The dataset is biased towards good conditions, and does not contain rough scenarios, like stormy weather or rough waters.
- Daytime: There are no night-time recordings.
- Geographic location: The recording locations are geographically limited to two locations, both being a near-shore scenario.
Further, there are two limitations:
- No leisure vehicles are in the validation split.
- Annotation quality is limited by human-level recognition and image quality, which results in a "flickering" of annotations for some cases. Although annotations were performed with high quality standards, annotation errors cannot be ruled out.
Please refer to the paper for a deeper discussion.
Citation
If you use the dataset, please cite it as:
@InProceedings{SeaClips,
title={SeaClips: A Video Dataset for Maritime Object Detection.},
authors={Denk, Franziska and Rankl, Christian and Almouahed, Shaban and Moser, David and Sablatnig, Robert},
booktitle={Winter Applications of Computer Vision (WACV)},
year=2026
}
Dataset Card Contact
For questions, concerns, or requests regarding the dataset, you can contact us via the following email: [email protected]. Please use the tag [SeaClips] in the subject.
References
[1] MMTracking Contributors. MMTracking: OpenMMLab video perception toolbox and benchmark. https://github.com/open-mmlab/mmtracking. 2020.
- Downloads last month
- 77