Source-linked AI summary
NTIRE 2026 Rip Current Detection and Segmentation (RipDetSeg) Challenge Report
Andrei Dumitriu, Aakash Ralhan, Florin Miron, Florin Tatui, Radu Tudor Ionescu, Radu Timofte, Abdullah Naeem, Anav Katwal, Ayon Dey, Md Tamjidul Hoque, Asuka Shin, Hiroto Shirono, Kosuke Shigematsu, Gaurav Mahesh, Anjana Nanditha, Jiji CV, Akbarali Vakhitov, Sang-Chul Lee, Xinger Li, Chun'an Yu, Junhao Chen, Yang Yang, Gundluri Yuvateja Reddy, Harshitha Palaram, Gejalakshmi N, Jeevitha S, Jiachen Tu, Guoyi Xu, Yaoxin Jiang, Jiajia Liu, Yaokun Shi, Amitabh Tripathi, Modugumudi Mahesh, Santosh Kumar Vipparthi, Subrahmanyam Murala
TL;DR
RipDetSeg addresses the difficult and safety-critical problem of recognizing rip currents whose visual appearance varies across coastal conditions. The challenge extends RipVIS by evaluating detection and segmentation with a composite metric, and its results show strong performance from pretrained vision models while identifying scope for methods tailored to rip-current structure.
Problem
Rip currents are difficult to identify automatically because their diffuse visual cues vary across environmental conditions, viewpoints, and beach morphologies, despite the safety-critical need for timely recognition.
Method
The challenge extends RipVIS by evaluating both instance segmentation and object detection, using a composite score that combines F1 and recall-oriented F2 across IoU thresholds.
Results
The top-performing method achieved computed scores of 55.79 for segmentation and 56.33 for detection, while submissions commonly used pretrained models with augmentation and post-processing.
Takeaways & Limitations
Rip-current understanding benefits from robust general-purpose vision models, while challenge results support continued development of methods tailored to the currents’ unique visual structure.
Takeaways & Limitations
Future editions may need broader architectures, temporal video information, and multimodal or contextual signals relevant to rip-current formation.
Abstract
from arXiv · showhide
This report presents the NTIRE 2026 Rip Current Detection and Segmentation (RipDetSeg) Challenge, which targets automatic rip current understanding in images. Rip currents are hazardous nearshore flows that cause many beach-related fatalities worldwide, yet remain difficult to identify because their visual appearance varies substantially across beaches, viewpoints, and sea states. To advance research on this safety-critical problem, the challenge builds on the RipVIS benchmark, evaluating both detection and segmentation. The dataset is diverse, sourced from more than $10$ countries, with $4$ camera orientations and diverse beach and sea conditions. This report describes the dataset, challenge protocol, evaluation methodology, final results, and summarizes the main insights from the submitted methods. The challenge attracted $159$ registered participants and produced $9$ valid test submissions across the two tasks. Final rankings are based on a composite score that combines $F_1[50]$, $F_2[50]$, $F_1[40\!:\!95]$, and $F_2[40\!:\!95]$. Most participant solutions relied on pretrained models, combined with strong augmentation and post-processing design. These results suggest that rip current understanding benefits strongly from the robust general-purpose vision models' progress, while leaving ample room for future methods tailored to their unique visual structure.
1. Introduction
RipDetSeg extends rip-current understanding from segmentation to both instance segmentation and detection, addressing a safety-critical problem whose visual cues vary across environments and viewpoints. Its diverse benchmark and joint evaluation target models that can localize and delineate currents for practical beach-safety systems.
- Motivation: Rip currents are hazardous seaward flows that are difficult for untrained observers to recognize in time.Their danger is heightened because swimmers may try to counter the flow rather than escape parallel to shore.
- Motivation: Automatic rip-current analysis remains challenging because currents have diffuse boundaries and subtle, unstable cues that vary with conditions, viewpoints, and beach morphology.Wave disruptions, sediment plumes, and water-color changes can be confounded by lighting, weather, and perspective.
- Challenge scope: RipDetSeg extends RipVIS and RipSeg by evaluating both instance segmentation and object detection in still images.The two tasks respectively support precise spatial delineation and practical coarse localization.
- Challenge scope: The benchmark covers diverse environmental conditions, camera viewpoints, and coastal morphologies, using separated training, validation, and held-out test sets.Strict data separation is intended to support fair evaluation and discourage overfitting.
- Evaluation rationale: Joint detection and segmentation evaluation balances complementary accuracy-efficiency trade-offs relevant to beach-safety applications.The challenge uses a custom metric that emphasizes robustness and recall in safety-critical scenarios.
2. Challenge format and ranking
The challenge evaluates detection and segmentation through staged data releases and reproducibility checks, then ranks submissions with a composite metric combining precision-oriented F1 and recall-oriented F2 across IoU thresholds.
- Challenge protocol: The challenge comprises detection and segmentation tasks conducted through intro, validation, and test phases on CodaBench.Validation provided annotated training images and additional unlabeled images for prediction-format verification.
- Challenge protocol: Only teams passing the reproducibility test were included in the final rankings.The rankings were reported separately for segmentation and detection.
- Ranking metric: False negatives receive special attention because missed rip currents can potentially be deadly, motivating inclusion of the recall-oriented F2 score.False positives are described as nuisances, whereas false negatives carry greater safety consequences.
- Ranking metric: The composite score combines F1[50], F1[40:95], F2[50], and F2[40:95] across IoU thresholds.The metrics use IoU 0.50, while [40:95] averages thresholds from 0.40 through 0.95 in steps of 0.05.
- Example submission pipeline: Team UNO Pixel Pros’s segmentation pipeline averages predictions from six models before thresholding and converting connected regions into filtered instance masks.Five models use cross-validation and one uses the full training dataset; the threshold is 0.5 and the minimum area is 100 pixels.
3. Methods
Submitted methods used pretrained segmentation and detection models with ensembles, augmentation, and post-processing to produce task outputs. The reported pipelines span semantic-mask conversion, dual-branch fusion, and segmentation-first prediction.
- Team UNO Pixel Pros: Team UNO Pixel Pros used a pretrained SegFormer-B5 with a six-model ensemble for segmentation.The ensemble contained five cross-validation models and one model trained on the full dataset.
- Cross-team patterns: The submitted approaches illustrate a broader reliance on pretrained models, augmentation, and inference-time post-processing.These design choices appear across ensemble, fusion, and mask-conversion pipelines.
- Team SiGMoid: Team SiGMoid combined YOLO11m and YOLO11x detection-segmentation branches through deterministic fusion and refinement.Its pipeline included score fusion, weighted box fusion, segmentation refinement, and area filtering.
- Team Riposte: Team Riposte used segmentation labels as its only training signal and derived both segmentation and detection outputs from predicted masks.Bounding-box labels were intentionally withheld during training.
- Team UNO Pixel Pros: UNO converted binary foreground masks into instances by pixel-wise averaging, thresholding at 0.5, connected-components analysis, and removal of regions smaller than 100 pixels.The resulting instances were converted to COCO polygon format for submission.
3.2. Team SiGMoid
Team SiGMoid used a two-branch YOLO11 ensemble with separate detection and segmentation models, then fused and refined predictions through deterministic post-processing.
- Model design: Two branches combined YOLO11m and YOLO11x models, with separately trained detection and segmentation components.The recover branch used YOLO11m, while the second used YOLO11x; both used AdamW, cosine learning-rate scheduling, and shared augmentations.
- Inference: Inference generated predictions without test-time augmentation using low initial confidence thresholds of 0.03 for segmentation and 0.01 for detection.Predictions were produced at fixed scales before polygon refinement and score fusion.
- Post-processing: Segmentation polygons were rasterized, morphologically filtered, reduced to their largest contour, cleaned of regions below 24 pixels, and simplified.The pipeline used one closing iteration with a kernel size of 5 and polygon approximation epsilon ratio 0.001.
- Score fusion: Each segmentation prediction was matched to its single best detection box by IoU, without one-to-one assignment.When the best overlap exceeded 0.20, the segmentation score was updated; otherwise, it was reduced by a factor of 0.95.
- Final fusion: Final detection boxes used weighted box fusion at IoU 0.45 across raw detections and boxes derived from refined segmentation predictions.Final filtering used confidence thresholds of 0.16 for segmentation and 0.18 for detection.
3.3. Team Riposte
Team Riposte used a lightweight segmentation-first YOLOv8n-seg system, training only with polygon annotations and deriving detection boxes from predicted masks.
- Method: The method trained only a YOLOv8n-seg segmentation model using polygon annotations, intentionally excluding bounding-box labels.Detection outputs were obtained at inference by computing axis-aligned boxes from predicted polygons.
- Method: Predicted masks served both tasks: instance masks supported segmentation, while their derived boxes supported detection.This segmentation-first design avoided separate detection and segmentation models.
- Training: Training used 20 epochs at 640×640 resolution with batch size 16 and Adam, retaining mosaic, horizontal flipping, and HSV jitter augmentations.The configuration used default YOLOv8 hyperparameters.
- Related submission: Team SoloSeg adopted a standard end-to-end YOLOv8s-seg setup initialized from pretrained COCO weights, without an ensemble or additional post-processing branches.The system selected checkpoints and thresholds using validation data.
3.5. Team KMG
Team KMG used a two-stage pipeline that ensembles YOLOv13 and DINOv3-based detection before refining Mask2Former segmentation with detection results.
- Pipeline: Detection was performed first and used to refine subsequent segmentation results.The pipeline explicitly separated detection and segmentation into two stages.
- Detection: KMG ensembled a YOLOv13 detector and a high-resolution DINOv3-based detector using IoU- and IoA-based merging.Matched detector boxes were merged by averaging coordinates and combining confidence scores, while unmatched high-confidence boxes were retained.
- Segmentation: Mask2Former with a DINOv3 backbone generated instance masks that were refined through geometric cleaning and splitting.The first refinement stage removed small or noisy fragments and simplified polygon shapes.
- Segmentation: A second detection-guided refinement stage filtered and aligned segmentation masks using the final detection results.This connected the segmentation output to the preceding detection stage.
3.6. Team RIP YuvatejaReddy
Team RIP YuvatejaReddy used a single COCO-initialized YOLOv8s-seg model to jointly predict detection boxes and segmentation polygons.
- Model: A single YOLOv8s-seg model jointly performed detection and segmentation.The model output both bounding-box coordinates and corresponding polygons.
- Training: Training used COCO-pretrained weights, a 90/10 validation split, 33 epochs at 640×640, batch size 4, and early stopping with patience 10.Default YOLO hyperparameters were used.
3.7. Team VisionX
Team VisionX made segmentation the primary prediction and derived detection boxes from mask extents, then used multi-scale fusion and morphological refinement to improve consistency and boundaries.
- VisionX used a YOLOv8s-seg model without explicit detection supervision, deriving bounding boxes directly from mask extents.This design enforces spatial alignment and avoids inconsistencies from separate box regression.
- Multi-scale inference aggregated augmented views, followed by overlap-based mask fusion and morphological refinement.The refinement removes noise and improves boundary quality.
- The mask-centric approach tightly couples segmentation and detection, making mask quality the basis for both tasks.
3.8. Team NTR
Team NTR combined independent YOLO11x detection and segmentation models, fusing boxes with WBF and using high-resolution, low-threshold inference with mask post-processing to emphasize recall.
- NTR used a two-model YOLO11x ensemble with separate detection and segmentation models, fusing their boxes via WBF.Segmentation masks came only from the YOLO11x-seg model.
- Both models ran at 1536×1536 with a 0.003 confidence threshold and NMS IoU 0.6 to maximize recall.Bounding boxes were fused with WBF using IoU 0.5 and equal weights.
- Segmentation masks were post-processed with morphological opening using a 5 × 5 kernel for 3 iterations.
- The approach combined high-resolution inference, aggressive recall tuning, and model ensembling to improve detection and segmentation performance.
3.9. Team Amitabh
Team Amitabh used a single YOLO11s segmentation model to jointly predict bounding boxes and instance masks with standard inference and no ensembling or post-processing.
- Amitabh used one YOLO11s segmentation model to jointly predict bounding boxes and instance masks.
- The model was trained on an 85/15 train-validation split for 200 epochs at input resolution 1024 with batch size 8.
- Inference used a 0.05 confidence threshold and NMS IoU threshold of 0.7, without ensembling or post-processing.
4. Conclusion
The challenge evaluated nine submissions for rip current detection and segmentation using four F1/F2 metrics, with results showing strong reliance on pretrained YOLO-based models and substantial remaining difficulty.
- The challenge evaluated nine final submissions across F1[50], F1[40:95], F2[50], and F2[40:95] for detection and instance segmentation.
- Most solutions relied on pretrained detection and segmentation models, especially YOLO-based instance segmentation, with augmentation and fine-tuning.
- Inference-time design, including ensembling, mask refinement, and geometry-aware post-processing, often drove gains in spatial consistency.
- 55.79 was the top computed score for segmentation, while 56.33 was the top computed score for detection.
- Future editions may examine broader architectures, video temporal information, multimodal data, and contextual signals about rip current formation.