AIM 2025 RipSeg Challenge

• AIM 2025 RipSeg Challenge is a benchmark for automatic rip current segmentation in coastal images, emphasizing accurate instance masks.
• It adapts the extensive RipVIS dataset with COCO-style polygon annotations and diverse camera viewpoints to highlight real-world complexities.
• The challenge drives innovations in domain adaptation and synthetic augmentation to accurately capture amorphous rip current features.

The AIM 2025 Rip Current Segmentation (RipSeg) Challenge is a benchmark competition for automatic rip current segmentation in still images. It was introduced to advance computer-vision methods for a safety-critical coastal hazard whose visual manifestation is amorphous, lacks rigid boundaries, and depends on subtle cues such as disrupted wave-breaking patterns, sediment plumes, localized changes in water color, and deflection flows. The challenge is built on RipVIS, described as the largest publicly available rip current dataset for video instance segmentation, but reformulates the problem as single-class instance segmentation on still images so that methods must delineate the full extent of each rip current rather than rely on coarse boxes or image-level labels (Dumitriu et al., 18 Aug 2025, Dumitriu et al., 1 Apr 2025).

1. Hazard model and task motivation

Rip currents are powerful, fast-moving surface flows that pull water seaward from the shore. The challenge report notes peak speeds exceeding 8.7 km/h, faster than an Olympic swimmer, and frames accurate visual detection as important for proactive warnings, lifeguard operations, and public safety campaigns (Dumitriu et al., 18 Aug 2025).

From a CV standpoint, the task is difficult because rip currents are not compact objects with stable contours. Their visible signatures vary with hydrodynamics, seabed morphology, shoreline structures, camera viewpoint, lighting, and weather. These factors induce substantial domain shift across scenes. The challenge therefore centers on instance segmentation rather than classification or bounding-box detection: precise masks are preferred because they delineate the full extent of the rip current and reduce irrelevant background (Dumitriu et al., 18 Aug 2025).

This design choice also aligns with the underlying RipVIS benchmark, which treats rip currents as instance-level phenomena that may appear, evolve, split, or fade over time. RipVIS emphasizes polygonal masks rather than coarse boxes, and it defines rip regions through visible flow channels characterized by gaps in wave breaking, offshore sediment transport plumes, or boundary-controlled currents adjacent to jetties, piers, or deflected breaker lines (Dumitriu et al., 1 Apr 2025).

2. Dataset basis, image split, and annotation protocol

RipSeg is built on RipVIS and adapts that video benchmark to still-image instance segmentation. The report states that RipVIS is the largest publicly available rip current dataset for video instance segmentation, and the challenge also includes training data from Dumitriu et al. (CVPRW 2023). For the image challenge, the dataset contains 27,718 images, divided into training, validation, and test partitions with COCO-style polygon annotations for a single target class, “rip current,” while allowing multiple instances per image (Dumitriu et al., 18 Aug 2025).

Split	Images	Annotation and evaluation status
Training	18,386	Polygon annotations in COCO JSON
Validation	4,348	COCO JSON without annotations; ground truth private on server
Test	4,984	No annotations; randomized filenames/hashes

The still-image corpus spans four camera orientations: aerial bird’s-eye, aerial tilted, elevated beachfront, and water-level beachfront. This viewpoint diversity is a central source of difficulty because feature distributions differ strongly across these acquisition modes. The report emphasizes diversity in locations, rip-current types, and viewpoints, but it does not enumerate specific countries, sites, or type counts for the challenge split itself. Resolution ranges and instance counts per image are also not specified (Dumitriu et al., 18 Aug 2025).

RipVIS provides the broader provenance for this design. It contains 184 videos and 212,328 frames, including 150 videos with rip currents and 34 videos without rip currents, collected from drones, mobile phones, fixed beach cameras, Internet sources, and prior data releases. Its geographic coverage includes USA, Mexico, Costa Rica, Portugal, Italy, Greece, Romania, Sri Lanka, Australia, and New Zealand, and most videos are annotated at 5 FPS. Manual masks were created on 15,784 frames and interpolated to produce 163,528 labeled frames containing rips (Dumitriu et al., 1 Apr 2025).

The challenge’s data handling included several anti-fraud and transparency measures. Evaluation code was publicly available from the beginning. During validation, file naming conventions indicated the presence or absence of rip currents, whereas final test filenames were randomized to prevent manual labeling (Dumitriu et al., 18 Aug 2025).

3. Competition framework, submission protocol, and ranking metric

The challenge was hosted on the Codabench evaluation server and was organized into three phases: a training phase of approximately two weeks, a validation phase of approximately five weeks, and a two-day test phase. Participants received 18,386 annotated training images, then 4,348 validation images with private ground truth on the server, and finally 4,984 blind test images. The test phase allowed a single submission, described as “only one try.” After submission, teams were required to provide a team description, reproducible code, and an approach description, and only teams passing this reproducibility check were included in the final ranking. In total, 75 participants registered and 5 valid test submissions were accepted (Dumitriu et al., 18 Aug 2025).

Evaluation used standard instance-segmentation measures. For a prediction mask $P$ and ground-truth mask $G$,

$\mathrm{IoU} = \frac{|P \cap G|}{|P \cup G|}.$

Precision and recall were defined as

$$P = \frac{TP}{TP + FP}, \qquad R = \frac{TP}{TP + FN}.$$

The challenge then computed

$F_1 = \frac{2PR}{P + R},$

and

$$F_2 = \frac{(1 + 2^2)PR}{2^2P + R} = \frac{5PR}{4P + R}.$$

Average Precision was measured both at a fixed IoU threshold and across a range of thresholds:

• $AP_{50}$: AP at IoU threshold $0.50$.
• $AP_{[50:95]}$: mean AP across IoU thresholds from $0.50$ to $G$0 in steps of $G$1.

Final ranking was determined by a weighted composite score,

$G$2

The report states that this weighting was chosen to mitigate the safety-driven bias of $G$3, which can be increased by raising false positives. The composite metric therefore balances localization quality, represented by $G$4 and $G$5, with detection accuracy and recall, represented by $G$6 and $G$7 (Dumitriu et al., 18 Aug 2025).

A recurrent misconception in safety-oriented detection is that maximizing $G$8 alone is sufficient. RipSeg explicitly rejects that premise: high recall remains important, but naive recall boosting can increase false positives, which is why the final leaderboard was not based on $G$9 alone (Dumitriu et al., 18 Aug 2025).

4. Methodological patterns in the submitted systems

The top submissions illustrate three main directions identified in the report: domain adaptation and domain generalization, one-to-one instance segmentation, and collaborative fusion with strong augmentation. Across the final leaderboard, the strongest methods leveraged deep learning architectures, pretrained models, domain adaptation techniques, domain generalization strategies, and synthetic augmentation to improve robustness under diverse conditions (Dumitriu et al., 18 Aug 2025).

RipEye, ranked first, used HRDA, an unsupervised domain-adaptive semantic segmentation framework, together with SHADE for style-hallucinated domain generalization. The team reported a significant feature gap between training and validation sets and selected domain generalization to improve robustness to unseen viewpoint and weather conditions. A notable architectural decision was the removal of SHADE’s retrospection consistency module because the pretraining dataset lacked rip current examples. RipEye also used RipGAN to generate 5,141 synthetic images from polygons and bounding boxes sampled from 50% of rip-current training images, thereby increasing style diversity. Confidence scoring was obtained by subtracting background probability from rip-current probability. Training used AdamW with $\mathrm{IoU} = \frac{|P \cap G|}{|P \cup G|}.$0, $\mathrm{IoU} = \frac{|P \cap G|}{|P \cup G|}.$1, weight decay $\mathrm{IoU} = \frac{|P \cap G|}{|P \cup G|}.$2, encoder learning rate $\mathrm{IoU} = \frac{|P \cap G|}{|P \cup G|}.$3, decoder learning rate $\mathrm{IoU} = \frac{|P \cap G|}{|P \cup G|}.$4, and a 1,500-iteration linear warm-up followed by linear decay. HRDA settings included random cropping at $\mathrm{IoU} = \frac{|P \cap G|}{|P \cup G|}.$5, flipping, a high-resolution detail crop of $\mathrm{IoU} = \frac{|P \cap G|}{|P \cup G|}.$6, and 98 basic styles. At inference, the team used only low-resolution context cropping and reported no significant performance drop. Compute was an NVIDIA RTX 4090 with 24 GB VRAM (Dumitriu et al., 18 Aug 2025).

RipSense, ranked second, used SparseInst with a PVTv2-B1 backbone. Its central methodological claim was that one-to-one instance segmentation via bipartite matching and instance activation maps reduces duplicate predictions on amorphous regions and therefore reduces dependence on NMS. The team added morphological opening, defined as erosion followed by dilation, with kernel size 5 and 3 iterations to refine and smooth shapes. It reported that CNN detectors such as YOLO11 variants tended to produce overlapping predictions for the same amorphous region; NMS at $\mathrm{IoU} = \frac{|P \cap G|}{|P \cup G|}.$7–$\mathrm{IoU} = \frac{|P \cap G|}{|P \cup G|}.$8 was often ineffective, while lowering NMS to $\mathrm{IoU} = \frac{|P \cap G|}{|P \cup G|}.$9–$$P = \frac{TP}{TP + FP}, \qquad R = \frac{TP}{TP + FN}.$$0 could suppress true positives when multiple rip currents were close together. Training used 10 epochs, AdamW with learning rate $$P = \frac{TP}{TP + FP}, \qquad R = \frac{TP}{TP + FN}.$$1, $$P = \frac{TP}{TP + FP}, \qquad R = \frac{TP}{TP + FN}.$$2, $$P = \frac{TP}{TP + FP}, \qquad R = \frac{TP}{TP + FN}.$$3, $$P = \frac{TP}{TP + FP}, \qquad R = \frac{TP}{TP + FN}.$$4, weight decay $$P = \frac{TP}{TP + FP}, \qquad R = \frac{TP}{TP + FN}.$$5, WarmupMultiStepLR with linear warmup for the first 400 iterations at factor $$P = \frac{TP}{TP + FP}, \qquad R = \frac{TP}{TP + FN}.$$6, and decays at epochs 6 and 8 by $$P = \frac{TP}{TP + FP}, \qquad R = \frac{TP}{TP + FN}.$$7. Batch size was 4, augmentation was horizontal flip, and inference used a confidence threshold of 0.4 on an NVIDIA RTX 3060 (Dumitriu et al., 18 Aug 2025).

Gogogochufalou, ranked third, implemented a collaborative dual-model framework combining SparseInst with a PVTv2-B2-li backbone and RTMDet initialized from RTMDet-Ins-x and fine-tuned end-to-end. The system predicted on unlabeled data, treated low-confidence samples as blank backgrounds, and generated new images by pasting rip-current instances from the training set through Simple Copy-Paste, with at most three pasted instances per image. Additional augmentation included rotations, scaling, flipping, color jitter, and multi-scale training for RTMDet. During fusion, IoU and Dice thresholds were used to identify matching detections, and the higher-confidence prediction was retained. SparseInst used base learning rate 0.0000125, batch size 4, and maximum resolution $$P = \frac{TP}{TP + FP}, \qquad R = \frac{TP}{TP + FN}.$$8. Compute was a mixed GPU cluster with 4 RTX 2080 Ti on one node and 2 RTX 3090 on another (Dumitriu et al., 18 Aug 2025).

ZYS, ranked fourth, selected YOLO11x for instance segmentation, emphasizing its deep architecture and optimized feature fusion for multi-scale features, as well as speed and precision for potential real-time monitoring. Training ran for 45 epochs with initial learning rate 0.001 decayed by a factor of 0.1 every 10 epochs. Augmentations were random horizontal flip with $$P = \frac{TP}{TP + FP}, \qquad R = \frac{TP}{TP + FN}.$$9, brightness and contrast adjustment of $F_1 = \frac{2PR}{P + R},$0, and random cropping from $F_1 = \frac{2PR}{P + R},$1 to $F_1 = \frac{2PR}{P + R},$2 original size. Inference used confidence threshold 0.15, NMS IoU threshold 0.6, and a maximum of 300 detections (Dumitriu et al., 18 Aug 2025).

Simplehh, ranked fifth, used a lightweight YOLOv8n model fine-tuned for the task and inserted four CBAM blocks after the Neck’s C2f block to enhance discriminability between foreground and background while maintaining low latency. The rationale was that video frames in the dataset are highly redundant because adjacent frames are similar, so a lightweight model may reduce overfitting risk, while CBAM may improve focus on rip regions. Training used AdamW with $F_1 = \frac{2PR}{P + R},$3, $F_1 = \frac{2PR}{P + R},$4, learning rate decayed from 0.01 to 0.001, image size $F_1 = \frac{2PR}{P + R},$5, batch size 64, and 50 epochs. Reported losses were binary cross-entropy, complete IoU loss, dice loss, and cross-entropy loss. Augmentations included HSV adjustment, random cropping, translation, horizontal flipping, random erasing, and mix-up. Compute was a single vGPU-32 device with PyTorch 2.7.1 (Dumitriu et al., 18 Aug 2025).

5. Final leaderboard, performance interpretation, and failure modes

The final test leaderboard reflects a narrow performance range at the top and a marked drop in localization quality for lower-ranked entries. The report’s final ranking is as follows (Dumitriu et al., 18 Aug 2025):

Rank	Team	Final Score
1	RipEye	0.68
2	RipSense	0.65
3	Gogogochufalou	0.64
4	ZYS	0.55
5	Simplehh	0.54

The detailed metrics were: RipEye (username: shenyang115), $F_1 = \frac{2PR}{P + R},$6, $F_1 = \frac{2PR}{P + R},$7, $F_1 = \frac{2PR}{P + R},$8, $F_1 = \frac{2PR}{P + R},$9; RipSense (username: aakash), $$F_2 = \frac{(1 + 2^2)PR}{2^2P + R} = \frac{5PR}{4P + R}.$$0, $$F_2 = \frac{(1 + 2^2)PR}{2^2P + R} = \frac{5PR}{4P + R}.$$1, $$F_2 = \frac{(1 + 2^2)PR}{2^2P + R} = \frac{5PR}{4P + R}.$$2, $$F_2 = \frac{(1 + 2^2)PR}{2^2P + R} = \frac{5PR}{4P + R}.$$3; Gogogochufalou (username: luopuu), $$F_2 = \frac{(1 + 2^2)PR}{2^2P + R} = \frac{5PR}{4P + R}.$$4, $$F_2 = \frac{(1 + 2^2)PR}{2^2P + R} = \frac{5PR}{4P + R}.$$5, $$F_2 = \frac{(1 + 2^2)PR}{2^2P + R} = \frac{5PR}{4P + R}.$$6, $$F_2 = \frac{(1 + 2^2)PR}{2^2P + R} = \frac{5PR}{4P + R}.$$7; ZYS (username: zhangcc), $$F_2 = \frac{(1 + 2^2)PR}{2^2P + R} = \frac{5PR}{4P + R}.$$8, $$F_2 = \frac{(1 + 2^2)PR}{2^2P + R} = \frac{5PR}{4P + R}.$$9, $AP_{50}$0, $AP_{50}$1; and Simplehh (username: gl0ria), $AP_{50}$2, $AP_{50}$3, $AP_{50}$4, $AP_{50}$5. The report explicitly notes that even the top score of 0.68 indicates that the task remains difficult under diverse conditions (Dumitriu et al., 18 Aug 2025).

Several analytic points emerge from the submitted systems. First, one-to-one instance segmentation, exemplified by SparseInst, helped avoid duplicate masks that often affect NMS-based pipelines on amorphous targets. Second, domain generalization and domain adaptation, especially the HRDA plus SHADE configuration, improved robustness across viewpoints and weather. Third, synthetic augmentation via RipGAN and copy-paste strategies broadened style coverage and increased data diversity. Fourth, morphological opening refined mask boundaries for irregular shapes and reduced spurious fragments. Fifth, multi-scale training improved handling of rip currents of different sizes, while IoU- and Dice-based fusion could preserve confident predictions across complementary models (Dumitriu et al., 18 Aug 2025).

The report also identifies characteristic failure modes. Domain gaps between training and validation or test data caused mislocalization and missed detections. CNN-based detectors often produced multiple overlapping masks for irregular currents, and NMS tuning could remove true positives when neighboring rips were close together. Scenes in which currents blended into the background, including sediment-driven currents, required broader spatial context; transformer-based or otherwise context-rich models performed relatively better in those cases. The subtle visual cues that signal rip currents were frequently obscured by perspective, lighting, or weather (Dumitriu et al., 18 Aug 2025).

These observations are consistent with the broader RipVIS benchmark, where SparseInst variants achieved the highest recall and $AP_{50}$6, YOLO11-l achieved the highest precision, and Temporal Confidence Aggregation improved recall and $AP_{50}$7 across models in video settings. RipVIS also reported that sediments, foam patches, and rapid camera motion could degrade performance or induce false positives, underscoring that the image challenge inherits a broader problem of instability across acquisition conditions (Dumitriu et al., 1 Apr 2025).

6. Reproducibility, limitations, and prospective research directions

Reproducibility was built into the challenge protocol. Evaluation ran on Codabench, evaluation code was public from the beginning, and post-test inclusion in the leaderboard required reproducible code, team details, and an approach description. Training annotations were distributed as COCO JSON polygons for single-class instance segmentation. The report also points to the AIM 2025 workshop page at https://cvlai.net/aim/2025/, while specific public team code links were not included (Dumitriu et al., 18 Aug 2025).

Several operational details remain explicitly unspecified in the report. It does not provide instance counts per image, resolution ranges, precise geographic sites for the challenge split, or strict rule constraints on external data or pretrained models. The report also does not specify runtime or resource limits. Because several top methods used pretrained architectures and synthetic augmentation, a plausible implication is that these practices were permitted, but this remains an inference rather than a formal rule statement (Dumitriu et al., 18 Aug 2025).

The organizers distill three principal lessons. Rip current instance segmentation is safety-critical and intrinsically hard; high $AP_{50}$8 alone is insufficient because false positives can rise when recall is increased naively. Domain adaptation, domain generalization, and synthetic augmentation are effective for bridging feature gaps across viewpoints and environmental conditions. One-to-one instance segmentation and careful post-processing, including morphological opening, are useful for managing amorphous shapes and duplicate masks (Dumitriu et al., 18 Aug 2025).

Future directions proposed in the report extend beyond current paradigms. They include integrating temporal information from video, which is RipVIS’s native modality; incorporating multi-modal sensing such as environmental or hydrodynamic signals; using contextual cues that influence rip formation; expanding the dataset to more locations and conditions; sharing strong open baselines; and refining evaluation frameworks so that they better reflect operational beach-safety needs (Dumitriu et al., 18 Aug 2025). In the context of RipVIS, these directions are compatible with temporal post-processing such as Temporal Confidence Aggregation, tracking-based instance association, and other video-native methods that improve stability across frames (Dumitriu et al., 1 Apr 2025).