PaperlayerView exact revision on arXiv

Source-linked AI summary

A Dataset with Multibeam Forward-Looking Sonar for Underwater Object Detection

Kaibing Xie, Jian Yang, Kang Qiu

arXiv:2212.00352v2cs.CVcs.AI

TL;DR

MFLS object-detection research has few public datasets, while collecting and annotating underwater sonar data requires substantial cost, time, labor, and expertise. UATD provides high-resolution raw MFLS images collected in lake and shallow water, annotated across 10 target categories, with Faster-RCNN and YOLOv3 benchmarks. Faster-RCNN with Resnet-18 achieved the best mAP of 83.9% and mAR of 89.7%, while YOLOv3 with MobilenetV2 reached 93.4 FPS.

Abstract

from arXiv · show

Multibeam forward-looking sonar (MFLS) plays an important role in underwater detection. There are several challenges to the research on underwater object detection with MFLS. Firstly, the research is lack of available dataset. Secondly, the sonar image, generally processed at pixel level and transformed to sector representation for the visual habits of human beings, is disadvantageous to the research in artificial intelligence (AI) areas. Towards these challenges, we present a novel dataset, the underwater acoustic target detection (UATD) dataset, consisting of over 9000 MFLS images captured using Tritech Gemini 1200ik sonar. Our dataset provides raw data of sonar images with annotation of 10 categories of target objects (cube, cylinder, tyres, etc). The data was collected from lake and shallow water. To verify the practicality of UATD, we apply the dataset to the state-of-the-art detectors and provide corresponding benchmarks for its accuracy and efficiency.

Background & Summary

MFLS supports detailed underwater sensing, but object-detection research is constrained by scarce public datasets and limitations in sonar image representations. UATD addresses these gaps with raw MFLS images, 10 target categories, and detector benchmarks.

Background & Summary

  • Earlier MFLS datasets had drawbacks including small samples, few object categories, or virtual images, and most were not public.The paper also describes prior sonar datasets for fish classification, marine-debris segmentation, and cross-view localization.Source: Background & Summary · pp. 1–2
  • Underwater MFLS datasets are scarce partly because collection and annotation require substantial cost, time, labor, and expertise.Simulated and translated sonar images differ from real images because of acoustic propagation complexity and underwater environmental instability.Source: Background & Summary · pp. 1–2
  • Human-oriented filtering and pseudo-coloring can lose effective data, while sector representation improves visual perception but adds invalid information beyond the sector.The paper contrasts raw and processed images containing the same three targets.Source: Background & Summary · p. 2, Figure 1 · p. 2
  • UATD provides high-resolution raw MFLS images from lake and shallow-water collection, annotations for 10 target categories, and accuracy and efficiency benchmarks.The authors state that the dataset aims to improve the availability of data for MFLS object-detection research.Source: Background & Summary · p. 2

Methods

The methods collect raw Gemini 1200ik sonar data across selected underwater objects and sites, then use stream-based annotation to label targets. The adjustable boat-mounted sonar setup and continuous playback support collection and annotation.

Methods

  • Gemini 1200ik sonar operates at 720 kHz for long-range detection and 1200 kHz for higher-resolution imaging at shorter ranges.The sonar's software development kit provides raw sonar image data for Windows and Linux.Source: Collecting MFLS Data · p. 2
  • The sonar is fixed to a box on a metal rod mounted to the boat hull, with an adjustable connection controlling sonar depth and tilt.The setup is shown in the data-collection equipment figure.Source: Collecting MFLS Data · pp. 2–3, Figure 3 · p. 3
  • The dataset covers 10 target categories, with each object suspended in water or placed on the bottom during collection.The objects were tethered individually to floating balls to help distinguish their rough locations; the boat-mounted sonar cruised around selected sites while recording.Source: Collecting MFLS Data · pp. 3–4
  • OpenSLT plays raw sonar data as a video stream so annotators can pause on targets and label them while tracking their appearances across frames.It also retains sonar operating information extracted from each frame and saves annotations locally for batch processing.Source: Object annotations · p. 4, Figure 6 · p. 4

Data Records

UATD contains 9,200 paired image and annotation files, with train and test partitions. Its records preserve sonar operating parameters, image information, and object boxes, while the dataset statistics describe class, range, and frequency distributions.

Data Records

  • 9,200 BMP images have matching XML annotations, split into 7,600 training pairs and two test sets of 800 pairs each.The files are organized into separate image and annotation folders within three ZIP archives.Source: Data Records · p. 5
  • The collection includes 2,900 images at 720 kHz and 6,300 at 1200 kHz, with sonar working ranges from 5 to 25 meters.The dataset statistics also show class distribution and counts by one-meter range intervals.Source: Data Records · p. 5, Figure 7 · p. 5
  • Each annotation records sonar parameters, file paths, image dimensions, and object categories with pixel bounding boxes.The example lists range, azimuth, elevation, sound speed, and frequency in the sonar section; the image example is 1024x1428 with three channels.Source: Data Records · p. 6

Technical Validation

The section validates UATD with object-detector benchmarks that report accuracy and efficiency, alongside a multi-step annotation-checking process. Faster-RCNN with Resnet-18 leads the reported accuracy metrics, while YOLOv3 with MobilenetV2 offers the strongest stated efficiency figures.

Technical Validation

  • Three team members annotated independently, cross-checked labels in OpenSLT against recorded video, and obtained a final check from a professional data-management company.The video stream was recorded synchronously with raw data to help detect and track objects during annotation.Source: Technical Validation · p. 6
  • The UATD benchmark evaluates Faster-RCNN and YOLOv3 with various backbones using MMdetection (V2.25.0).The authors describe these as popular two-stage and one-stage detectors, respectively.Source: Object detection benchmarks · p. 6
  • The evaluation measures accuracy with mAP at IoU 0.5 and mAR, and efficiency with FPS, Params, and FLOPs.Training and inference resize image height to 512 pixels while preserving the original width ratio; training uses 12 epochs on a GTX 1080 GPU.Source: Object detection benchmarks · p. 7
  • Faster-RCNN with Resnet-18 achieves the best mAP of 83.9% and mAR of 89.7%, while YOLOv3 with MobilenetV2 reaches 93.4 FPS.YOLOv3 with MobilenetV2 also has 3.68M Params and 4.22G FLOPs.Source: Object detection benchmarks · p. 7, Table 2 · p. 7

Code availability

The UATD dataset and OpenSLT annotation tool are available through a figshare repository, with a test example and README guidance.

Code availability

  • The UATD dataset is published in a figshare repository, alongside the OpenSLT annotation tool archived as UATD_OpenSLT.zip.OpenSLT is developed based on Qt 5.9.Source: Code availability · p. 7
  • OpenSLT worked well during annotation on Ubuntu 18.04 and 20.04.Source: Code availability · p. 7
  • The tool includes a small example dataset with sonar images and corresponding CSV files, plus a README.md user guide.Source: Code availability · p. 7
Loading 2212.00352v2…