Rip Current Segmentation: A Novel Benchmark and YOLOv8 Baseline Results

TL;DR

This work addresses the danger of rip currents by introducing a first-of-its-kind benchmark for rip current instance segmentation, comprising polygon-annotated training images and 17 drone videos for testing. It evaluates multiple YOLOv8 configurations as a baseline, finding that the YOLOv8-nano model delivers the best balance of accuracy and real-time performance, with a validation mAP50 of 88.94% and a test macro average of 81.21%. The dataset and code are publicly available to foster further research and practical safety tools. Overall, the study provides a concrete, portable baseline for rip current segmentation and highlights directions for leveraging temporal information and broader class support in future work.

Abstract

Rip currents are the leading cause of fatal accidents and injuries on many beaches worldwide, emphasizing the importance of automatically detecting these hazardous surface water currents. In this paper, we address a novel task: rip current instance segmentation. We introduce a comprehensive dataset containing $2,466$ images with newly created polygonal annotations for instance segmentation, used for training and validation. Additionally, we present a novel dataset comprising $17$ drone videos (comprising about $24K$ frames) captured at $30 FPS$, annotated with both polygons for instance segmentation and bounding boxes for object detection, employed for testing purposes. We train various versions of YOLOv8 for instance segmentation on static images and assess their performance on the test dataset (videos). The best results were achieved by the YOLOv8-nano model (runnable on a portable device), with an mAP50 of $88.94%$ on the validation dataset and $81.21%$ macro average on the test dataset. The results provide a baseline for future research in rip current segmentation. Our work contributes to the existing literature by introducing a detailed, annotated dataset, and training a deep learning model for instance segmentation of rip currents. The code, training details and the annotated dataset are made publicly available at https://github.com/Irikos/rip_currents.

Figures (4)

• Figure 1: An illustration of a visible rip current and its detection task, displayed from left to right: 1. The original photo, 2. The bounding box annotation from the YOLO-Rip dataset, 3. Our ground-truth annotation for instance segmentation, 4. Prediction using YOLOv8. This example highlights how bounding boxes may exclude relevant parts of the rip currents while also incorporating surrounding noise.
• Figure 2: Rip current illustration showing the direction of the current and how to swim in order to escape it. Source: https://www.noaa.gov/.
• Figure 3: Selected images from the training dataset with corresponding ground-truth annotations. Rip currents exhibit an amorphous nature, resulting in varying appearances even for the same rip current, depending on the snapshot's timing and surrounding conditions. The top and middle rows display rip currents that are easy and difficult to spot, respectively. The bottom row showcases images without rip currents in various settings.
• Figure 4: Example of results on testing dataset (using the nano model). On the first two rows we have the correct predictions, the true positives and the true negatives. On the last row we have the incorrect predictions, the false positives and the false negatives. Notice column (d) on the last row, where the model manages to correctly predict one rip current, but misses the second one.