SWDD: Sonar Wall Detection Dataset

This repository contains three side scan sonar datasets: SWDD: Sonar Wall Detection Dataset, which is part of the paper Knowledge Distillation in YOLOX-ViT for Side Scan Sonar Object Detection. SWDD-Validation is an extension version and part of the paper ROSAR: An Adversarial Re-Training Framework for Robust Side-Scan Sonar Object Detection. SWDD-Adversarial is another extended version and part of the paper ROSAR: An Adversarial Re-Training Framework for Robust Side-Scan Sonar Object Detection. SWDD Dataset SWDD has been recorded with a lightweight autonomous underwater vehicle (LAUV), operated by OceanScan-MST and carrying a Klein 3500 side scan sonar (SSS). The AUV has been deployed in the Porto de Leixões harbor following the harbor's walls while collecting SSS raw data. The SSS operated at a high frequency of 900kHz with a range of 75m for a total range of 150m from port to starboard. This setup produced a resolution of 4.168 pixels per line. The data were processed using Neptus software, transforming the raw data into waterfall images. The 216 images have been manually annotated with two different classes, wall and noWall. Data augmentation such as noise, flips, and combined noise-flip transformations have been applied, increasing the data to 864 images. Across the 864 images, there are a total of 2,616 labeled samples. To ensure robust training, with respect to the data quality, the original dataset has been mixed with the augmented images. The dataset is divided into 70% for training, 15% for validation, and 15% for testing. The authors chose to generate images with 500 lines, meaning the images have a resolution of 4.168x500. Finally, the images have been resized to 640 × 640 to use this data in specific computer vision algorithms. The dataset is annotated following the COCO annotation format. YOLOX and YOLOX-ViT have been trained and compared using the SWDD dataset. A 6-minute 57-second video from another survey was used for model comparison. From this video, 6243 frames have been extracted with its manually annotated ground truth. Thus, this dataset repository offers an SSS dataset, a 6-minute 57-second SSS video, and 6243 extracted frames from this video with its manually annotated ground truth. The Knowledge Distillation in YOLOX-ViT code using the SWDD dataset is publicly available at KD-YOLOX-ViT. For (re-)using/publishing SWDD-Validation, please include the following copyright text: SWDD is a public dataset collected with a Light Autonomous Underwater Vehicle by Oceanscan-MST, within the scope of the H2020 REMARO project. ⚠️The two following datasets (SWDD-Validation and SWDD-Adversarial) are parts of the paper ROSAR: An Adversarial Re-Training Framework for Robust Side-Scan Sonar Object Detection, submitted to the 2025 IEEE Symposium on Maritime Informatics & Robotics (MARIS 2025). The paper is still under review. ⚠️ SWDD-Validation The SWDD dataset has been recorded using the LAUV with the Klein 3500 side scan sonar deployed in the Porto de Leixões harbor following the harbor's walls. The SWDD-Validation comports three field-collected detected: SWDD-Clean, SWDD-Surface, and SWDD-Noisy. The SWDD-Clean dataset, which includes data from the same mission as the original SWDD dataset; the SWDD-Surface dataset, captured while the LAUV was on the surface during windy weather, featuring a non-straight wall and wave-induced variations; and the SWDD-Noisy dataset, collected under stormy conditions, where the SSS transducer intermittently exited the water, resulting in data loss represented by black lines in the images. The SWDD-Validation dataset provides three datasets collected under different weather and sonar setups aiming to study object detection models' robustness variation under different sonar and noise conditions. The metadata of the SWDD-Validation dataset is depicted in the following table. Dataset # Image # Bbox Freq. (kHz) Range (m) Resolution SWDD-Clean 148 248 900 50 4168 x 500 SWDD-Surface 98 153 900 75 6552 x 500 SWDD-Noisy 551 800 455 100 4168 x 500 For (re-)using/publishing SWDD-Validation, please include the following copyright text: SWDD-Validation is a public dataset collected with a Light Autonomous Underwater Vehicle by Oceanscan-MST, within the scope of the H2020 REMARO project. SWDD-Adversarial The SWDD-Adversarial dataset has been generated by the ROSAR framework proposed in the paper ROSAR: An Adversarial Re-Training Framework for Robust Side-Scan Sonar Object Detection, which, by leveraging adversarial PGD and Patch attack on the SWDD dataset, generates the three following datasets: P1-SWDD, P2-SWDD, and Patch-SWDD. The ROSAR framework is publicly available in our GitHub [repository]. P1 and P2-SWDD datasets result from adversarial PGD attacks, each resulting from different safety properties (P1 and P2), resulting in 1017 and 1462 images for P1 and P2. Patch-SWDD dataset results from the adversarial patch attack, resulting in 151 images provided by the following repository: Patch Attack. For (re-)using/publishing SWDD-Adversarial, please include the following copyright text: SWDD-Adversarial is a public dataset based on the original SWDD dataset, which was collected using a Light Autonomous Underwater Vehicle by Oceanscan-MST and is within the scope of the H2020 REMARO project.本仓库包含三套侧扫声呐（Side Scan Sonar, SSS）数据集： SWDD：声呐墙体检测数据集（Sonar Wall Detection Dataset），为论文《YOLOX-ViT中的知识蒸馏用于侧扫声呐目标检测》的配套数据集。 SWDD-Validation为扩展版本，隶属于论文《ROSAR：面向鲁棒侧扫声呐目标检测的对抗重训练框架》。 SWDD-Adversarial为另一扩展版本，同样属于上述ROSAR论文。 ### SWDD数据集 SWDD由轻型自主水下航行器（Light Autonomous Underwater Vehicle, LAUV）采集，该航行器由OceanScan-MST运营，搭载Klein 3500型侧扫声呐。航行器在莱索昂港（Porto de Leixões）沿港壁部署，采集原始声呐数据。该声呐工作于900kHz高频，单测程75m，总覆盖范围从左舷到右舷达150m，每线分辨率为4.168像素。数据通过Neptus软件处理，将原始数据转换为瀑布图图像。原始数据集共216张图像，经人工标注为`wall`（墙体）与`noWall`（非墙体）两个类别。通过添加噪声、翻转、噪声-翻转组合等数据增强操作，将数据集扩充至864张图像，总计包含2616个标注样本。为保证训练鲁棒性，原始数据与增强数据混合后，按70%训练集、15%验证集、15%测试集划分。图像初始为500线分辨率，即尺寸为4.168×500，最终统一调整为640×640以适配特定计算机视觉算法。本数据集采用COCO标注格式。 研究团队使用SWDD数据集训练并对比了YOLOX与YOLOX-ViT模型。此外，使用另一项勘测的6分57秒视频开展模型对比，从该视频中提取6243帧并进行人工标注真值。 本数据集仓库提供一套侧扫声呐数据集、一段6分57秒的侧扫声呐视频，以及从该视频中提取的6243帧带人工标注真值的图像。 使用SWDD数据集的YOLOX-ViT知识蒸馏代码已在KD-YOLOX-ViT开源。 若需重新使用或发布SWDD-Validation，请包含以下版权声明： > SWDD是由Oceanscan-MST使用轻型自主水下航行器采集的公开数据集，隶属于H2020框架下的REMARO项目。 ⚠️以下两个数据集（SWDD-Validation与SWDD-Adversarial）属于已提交至2025年IEEE海事信息学与机器人研讨会（MARIS 2025）的论文《ROSAR：面向鲁棒侧扫声呐目标检测的对抗重训练框架》，该论文目前仍处于审稿阶段。⚠️ --- ### SWDD-Validation数据集 SWDD-Validation数据集同样由搭载Klein 3500侧扫声呐的轻型自主水下航行器在莱索昂港沿港壁采集。SWDD-Validation包含三个野外采集子集：SWDD-Clean、SWDD-Surface与SWDD-Noisy。其中SWDD-Clean与原始SWDD数据集来自同一航次；SWDD-Surface采集于航行器处于水面且有风的工况，包含非平直墙体与波浪引起的图像变化；SWDD-Noisy采集于风暴天气下，此时声呐换能器间歇性出水，导致图像中出现黑线代表的数据丢失情况。 SWDD-Validation数据集包含三种不同天气与声呐配置下采集的数据，用于研究目标检测模型在不同声呐工况与噪声条件下的鲁棒性变化。SWDD-Validation数据集的元数据如下表所示： | 数据集 | 图像数量 | 标注框数量 | 工作频率（kHz） | 测程（m） | 分辨率 | |----------------|----------|------------|----------------|----------|----------------------| | SWDD-Clean | 148 | 248 | 900 | 50 | 4168 × 500 | | SWDD-Surface | 98 | 153 | 900 | 75 | 6552 × 500 | | SWDD-Noisy | 551 | 800 | 455 | 100 | 4168 × 500 | 若需重新使用或发布SWDD-Validation，请包含以下版权声明： > SWDD-Validation是由Oceanscan-MST使用轻型自主水下航行器采集的公开数据集，隶属于H2020框架下的REMARO项目。 --- ### SWDD-Adversarial数据集 SWDD-Adversarial数据集由论文《ROSAR：面向鲁棒侧扫声呐目标检测的对抗重训练框架》中提出的ROSAR框架生成。该框架通过对SWDD数据集施加对抗PGD攻击与补丁攻击，生成以下三个子集：P1-SWDD、P2-SWDD与Patch-SWDD。ROSAR框架已在我们的GitHub仓库[repository]开源。 P1-SWDD与P2-SWDD数据集源自对抗PGD攻击，基于两种不同的安全属性（P1与P2）生成，分别包含1017张与1462张图像。 Patch-SWDD数据集源自对抗补丁攻击，包含151张图像，相关资源可在Patch Attack仓库获取。 若需重新使用或发布SWDD-Adversarial，请包含以下版权声明： > SWDD-Adversarial是基于原始SWDD数据集的公开数据集，原始SWDD数据集由Oceanscan-MST使用轻型自主水下航行器采集，隶属于H2020框架下的REMARO项目。