RipBench: A Unified Benchmark for Multi-Level Rip Current Detection, Classification and Segmentation
Organizations: Computer Vision Lab, CAIDAS & IFI, University of Würzburg, Germany · University of Bucharest, Romania
Abstract
Rip currents are a serious and often under-addressed threat to beach safety and the leading cause of coastal drownings worldwide. They are difficult to detect due to their amorphous structure, similarity to the background, and large variability in viewpoints and environments. Despite their dangers and growing interest in automated detection, existing research remains fragmented across datasets, metrics, models, and approaches, leading to an incomplete understanding of model performance. We introduce RipBench, a unified benchmark that enables controlled evaluation of rip current detection spanning multiple tasks, including classification, axis-aligned and oriented object detection, and instance, semantic, and panoptic segmentation, on the same data with standardized splits. This allows direct comparison of model performance on all levels of visual abstraction. Across videos ( frames, of them annotated for all six tasks) collected from diverse coastlines, our results expose a clear performance gap between coarse recognition and precise spatial understanding: frame-level classification reaches , while the best detector and the best instance segmenter reach and , with model rankings that change from task to task. Video-level confidence intervals, a blind inter-annotator study of the spatial labels, stratified analyses by viewpoint and visual signature, and video views with temporal baselines (single-seed, short-schedule reference runs) accompany the benchmark. All tasks are supported with carefully curated annotations and evaluated using standard metrics, with a focus on the score to emphasize recall in this safety-critical setting. RipBench, along with multiple baselines per task, is publicly available at https://RipBench.ai to support progress in real-world, multi-task vision for beach safety.
Figures & tables
| Detection | Segmentation | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Dataset | Videos | #Frames | #Labeled | Classif. | BBox | Or.BBox | Inst. | Sem. | Pan. |
| Maryan et al . (2019) [ 40 ] | ✗ | ✓ | ✗ | ✗ | ✗ | ✗ | |||
| de Silva et al . (2021) [ 20 ] | ✓ | ✗ | ✗ | ✗ | ✗ | ||||
| YOLO-Rip (2022) [ 61 ] | ✓ | ✗ | ✗ | ✗ | ✗ | ||||
| Dumitriu et al . (2023) [ 24 ] | ✓ | ✗ | ✗ | ||||||
| RipScout (2025) [ 34 ] | * | ✗ | ✓ | ✗ | ✗ | ✗ | ✗ | ||
| Model | Acc. | P | R | [95% CI] | AUROC | FPS | |
|---|---|---|---|---|---|---|---|
| ResNet-34 @640 | 0.875 | 0.847 | 0.890 | 0.868 | 0.881 [0.77, 0.94] | 0.960 | 108 |
| ResNet-50 @640 | 0.908 | 0.905 | 0.894 | 0.900 | 0.896 [0.78, 0.96] | 0.966 | 88 |
| ResNet-101 @640 | 0.900 | 0.908 | 0.871 | 0.889 | 0.878 [0.75, 0.95] | 0.959 | 80 |
| ResNet-152 @224 | 0.842 | 0.769 | 0.941 | 0.846 | 0.901 [0.78, 0.95] | 0.928 | 108 |
| YOLO26-n | 0.860 | 0.790 | 0.948 | 0.862 | 0.911 [0.81, 0.96] | 0.960 | 110 |
| YOLO26-x | 0.935 | 0.925 | 0.935 | 0.930 | 0.933 [0.85, 0.97] | 0.971 | 84 |
| Model | [95% CI] | FPS | ||||||
|---|---|---|---|---|---|---|---|---|
| Co-DETR-Swin-L ‡ | 0.683 | 0.150 | 0.268 | 0.601 | 0.303 | 0.692 [0.54, 0.80] | 0.348 | 14 |
| DINO-Swin-L ‡ | 0.609 | 0.071 | 0.211 | 0.655 | 0.293 | 0.679 [0.50, 0.80] | 0.303 | 7 |
| MaskDINO-Swin-L † | 0.588 | 0.174 | 0.257 | 0.638 | 0.352 | 0.612 [0.35, 0.78] | 0.337 | 17 |
| RF-DETR-n | 0.699 | 0.106 | 0.244 | 0.659 | 0.298 | 0.695 [0.51, 0.81] | 0.314 | 73 |
| RF-DETR-s | 0.718 | 0.132 | 0.280 | 0.681 | 0.323 | 0.715 [0.56, 0.83] | 0.340 | 70 |
| RF-DETR-m | 0.678 | 0.169 | 0.283 | 0.683 | 0.343 | 0.669 [0.48, 0.81] | 0.337 | 63 |
| Model | P | R | [95% CI] | FPS | ||
|---|---|---|---|---|---|---|
| YOLO26-n | 0.379 | 0.491 | 0.432 | 0.460 | 0.443 [0.16, 0.64] | 251 |
| YOLO26-s | 0.329 | 0.439 | 0.401 | 0.419 | 0.408 [0.15, 0.63] | 244 |
| YOLO26-m | 0.367 | 0.430 | 0.442 | 0.436 | 0.439 [0.18, 0.63] | 229 |
| YOLO26-l | 0.388 | 0.508 | 0.451 | 0.478 | 0.462 [0.23, 0.66] | 178 |
| YOLO26-x | 0.376 | 0.455 | 0.441 | 0.448 | 0.444 [0.18, 0.65] | 154 |
| Model | [95% CI] | FPS | ||||||
|---|---|---|---|---|---|---|---|---|
| Mask2Former-Swin-L | 0.563 | 0.072 | 0.206 | 0.630 | 0.283 | 0.630 [0.37, 0.79] | 0.283 | 23 |
| MaskDINO-Swin-L | 0.585 | 0.113 | 0.223 | 0.639 | 0.299 | 0.619 [0.36, 0.79] | 0.290 | 17 |
| RF-DETR-Seg-n | 0.651 | 0.128 | 0.258 | 0.671 | 0.324 | 0.656 [0.45, 0.81] | 0.317 | 54 |
| RF-DETR-Seg-s | 0.626 | 0.124 | 0.255 | 0.647 | 0.308 | 0.644 [0.42, 0.79] | 0.306 | 60 |
| RF-DETR-Seg-m | 0.612 | 0.112 | 0.249 | 0.606 | 0.296 | 0.632 [0.42, 0.77] | 0.308 | 50 |
| RF-DETR-Seg-l | 0.630 | 0.151 | 0.267 | 0.602 | 0.305 | 0.629 [0.40, 0.78] | 0.319 | 50 |
| Model | [95% CI] | Pixel acc. | FPS | |||||
|---|---|---|---|---|---|---|---|---|
| Mask2Former-R50 | 0.376 [0.18, 0.50] | 0.967 | 0.547 | 0.506 | 0.983 | 0.984 | 0.969 | 31 |
| MaskDINO-R50 | 0.397 [0.19, 0.52] | 0.969 | 0.568 | 0.519 | 0.984 | 0.986 | 0.972 | 31 |
| YOLO26-n | 0.406 [0.16, 0.55] | 0.955 | 0.577 | 0.508 | 0.977 | 0.978 | 0.964 | 423 |
| YOLO26-s | 0.342 [0.13, 0.48] | 0.958 | 0.510 | 0.436 | 0.978 | 0.980 | 0.965 | 426 |
| YOLO26-m | 0.379 [0.15, 0.51] | 0.965 | 0.550 | 0.468 | 0.982 | 0.984 | 0.970 | 362 |
| YOLO26-l | 0.395 [0.18, 0.53] | 0.964 | 0.567 | 0.501 | 0.982 | 0.983 | 0.969 | 265 |
| Model | [95% CI] | FPS | |||||
|---|---|---|---|---|---|---|---|
| Mask2Former-R50 | 0.396 [0.21, 0.52] | 0.696 | 0.568 | 0.943 | 0.990 | 0.806 | 41 |
| Mask2Former-Swin-L | 0.465 [0.27, 0.58] | 0.695 | 0.670 | 0.946 | 0.985 | 0.826 | 23 |
| MaskDINO-R50 | 0.433 [0.18, 0.58] | 0.709 | 0.611 | 0.954 | 0.996 | 0.824 | 30 |
| MaskDINO-Swin-L | 0.459 [0.25, 0.58] | 0.695 | 0.661 | 0.947 | 0.985 | 0.825 | 17 |
| Family | Configuration | [95% CI] | ||
|---|---|---|---|---|
| Per-frame image model | RF-DETR-Seg-x | 0.663 [0.46, 0.79] | 0.660 | 0.247 |
| MinVIS | MinVIS-R50 (dense-interp view) | 0.593 [0.29, 0.78] | 0.598 | 0.240 |
| CAVIS | CAVIS-Swin-L + offline refiner | 0.630 [0.36, 0.81] | 0.610 | 0.229 |
| Video-Mask2Former | Video-Mask2Former-R50 | 0.568 [0.31, 0.72] | 0.580 | 0.241 |
| DVIS++ (segmenter stage) | DVIS++-R50 | 0.517 [0.31, 0.65] | 0.464 | 0.180 |
Appendix figures & tables24 assets
Supplementary material from the paper’s appendix.
Appendix
| Comparison | Models (A B) | 95% CI | |
|---|---|---|---|
| cls-det (val-selected) | YOLO26-x @224 RF-DETR-n | +0.176 | [+0.048, +0.336] ∗ |
| cls-seg (val-selected) | YOLO26-x @224 Mask2Former-Swin-L | +0.241 | [+0.068, +0.454] ∗ |
| cls-obb (val-selected) | YOLO26-x @224 YOLO26-x | +0.427 | [+0.211, +0.677] ∗ |
| seg-det (val-selected) | Mask2Former-Swin-L RF-DETR-n | -0.065 | [-0.153, -0.006] ∗ |
| det-bfi (val-selected) | RF-DETR-n YOLO26-m | +0.178 | [+0.064, +0.328] ∗ |
| seg-det within RF-DETR (val-selected) | RF-DETR-Seg-l RF-DETR-n | -0.065 | [-0.142, -0.015] ∗ |
| Configuration | Seeds | Metrics: mean sd (values) |
| RT-DETR-l (bbox) | 0, 1, 2 | : 0.533 0.011 (0.535, 0.542, 0.520); : 0.467 0.028 (0.448, 0.499, 0.454); : 0.172 0.025 (0.153, 0.163, 0.200) |
| YOLO26-x (bbox) | 0, 1, 2 | : 0.515 0.015 (0.507, 0.505, 0.532); : 0.476 0.016 (0.479, 0.490, 0.459); : 0.197 0.013 (0.207, 0.200, 0.183) |
| ResNet-152 @224 (classification) | 0, 1, 2 | : 0.858 0.037 (0.901, 0.840, 0.834); AUROC: 0.907 0.029 (0.928, 0.920, 0.874); Acc.: 0.790 0.097 (0.842, 0.849, 0.678) |
| YOLO26-x (classification) | 0, 1, 2 | : 0.931 0.002 (0.933, 0.932, 0.930); AUROC: 0.964 0.011 (0.971, 0.952, 0.970); Acc.: 0.914 0.027 (0.935, 0.924, 0.885) |
| RF-DETR-Seg-n (instance segmentation) | 0, 1, 2 | : 0.647 0.008 (0.656, 0.642, 0.643); : 0.642 0.009 (0.651, 0.640, 0.634); : 0.258 0.008 (0.258, 0.267, 0.251) |
| Mask2Former-Swin-L (instance segmentation) | 0, 1, 2 | : 0.617 0.012 (0.630, 0.606, 0.616); : 0.553 0.012 (0.563, 0.540, 0.554); : 0.210 0.006 (0.206, 0.216, 0.207) |
| Task | Pairs | Median IoU | Mean IoU | Median Dice | ||
|---|---|---|---|---|---|---|
| Bounding boxes | 600 | 0.805 | 0.785 | 0.892 | 98.3% | 67.8% |
| Oriented boxes | 600 | 0.759 | 0.736 | 0.863 | 94.8% | 52.0% |
| Instance masks | 600 | 0.770 | 0.747 | 0.870 | 97.5% | 57.3% |
| Task | Bird’s eye | Aerial tilted | Elevated | Water-level |
|---|---|---|---|---|
| Bounding boxes | 0.794 (157) | 0.803 (273) | 0.813 (152) | 0.811 (18) |
| Oriented boxes | 0.750 (157) | 0.770 (273) | 0.745 (152) | 0.706 (18) |
| Instance masks | 0.751 (157) | 0.772 (273) | 0.777 (152) | 0.711 (18) |
| Train | Eval GT | nano | small | medium | large | xlarge |
|---|---|---|---|---|---|---|
| Tight (T) | tight | 0.444 / 0.409 / 0.191 | 0.425 / 0.351 / 0.136 | 0.504 / 0.374 / 0.151 | 0.505 / 0.430 / 0.184 | 0.507 / 0.479 / 0.207 |
| Tight (T) | derived | 0.409 / 0.373 / 0.137 | 0.407 / 0.327 / 0.106 | 0.482 / 0.345 / 0.111 | 0.446 / 0.354 / 0.124 | 0.467 / 0.400 / 0.116 |
| Derived (D) | tight | 0.495 / 0.448 / 0.179 | 0.506 / 0.441 / 0.181 | 0.529 / 0.450 / 0.167 | 0.471 / 0.452 / 0.166 | 0.521 / 0.489 / 0.179 |
| Derived (D) | derived | 0.437 / 0.352 / 0.162 | 0.489 / 0.401 / 0.154 | 0.517 / 0.437 / 0.153 | 0.467 / 0.458 / 0.228 | 0.502 / 0.473 / 0.233 |
| Signature (videos / instances) | Detection: RF-DETR-s | Inst. seg.: RF-DETR-Seg-x |
|---|---|---|
| dark gap (20 / 3608) | 0.829 / 0.856 | 0.788 / 0.821 |
| breaking wave gap (4 / 340) | 0.599 / 0.617 | 0.485 / 0.454 |
| sediment plume (6 / 2026) | 0.544 / 0.463 | 0.519 / 0.433 |
| foam plume (1 / 203) | 0.878 / 0.906 | 0.296 / 0.142 |
| Bucket (rip videos) | Detection: RF-DETR-s | Inst. seg.: RF-DETR-Seg-x |
|---|---|---|
| <720p (2) | 0.910 [0.86, 0.95] | 0.623 [0.53, 0.72] |
| 720p (11) | 0.569 [0.47, 0.79] | 0.487 [0.23, 0.71] |
| 1080p (9) | 0.824 [0.48, 0.88] | 0.852 [0.30, 0.91] |
| >=1440p (9) | 0.363 [0.06, 0.69] | 0.066 [0.00, 0.62] |
| Task (evaluation GT) | RipVIS-trained | RipBench-trained | Difference [95% CI] |
|---|---|---|---|
| Instance segmentation (instance masks) | 0.559 | 0.522 | +0.037 [-0.016, +0.136] |
| Derived boxes (mask-derived boxes) | 0.484 | 0.494 | -0.010 [-0.078, +0.052] |
| Model | Acc. | P | R | [95% CI] | AUROC | FPS | |
|---|---|---|---|---|---|---|---|
| ResNet-18 @224 | 0.840 | 0.840 | 0.808 | 0.824 | 0.814 [0.64, 0.90] | 0.903 | 226 |
| ResNet-18 @224 +add | 0.819 | 0.732 | 0.959 | 0.830 | 0.903 [0.81, 0.95] | 0.941 | 230 |
| ResNet-18 @640 | 0.766 | 0.681 | 0.924 | 0.784 | 0.863 [0.73, 0.93] | 0.930 | 120 |
| ResNet-18 @640 +add | 0.855 | 0.806 | 0.903 | 0.852 | 0.882 [0.78, 0.96] | 0.936 | 120 |
| ResNet-34 @224 | 0.881 | 0.902 | 0.832 | 0.866 | 0.845 [0.70, 0.93] | 0.933 | 204 |
| ResNet-34 @224 +add | 0.837 | 0.761 | 0.942 | 0.842 | 0.899 [0.79, 0.96] | 0.941 | 204 |
| Model | [95% CI] | FPS | ||||||
|---|---|---|---|---|---|---|---|---|
| Co-DETR-R50 +add ‡ | 0.583 | 0.160 | 0.231 | 0.617 | 0.305 | 0.613 [0.40, 0.75] | 0.302 | 24 |
| Co-DETR-Swin-L +add ‡ | 0.683 | 0.150 | 0.268 | 0.601 | 0.303 | 0.692 [0.54, 0.80] | 0.348 | 14 |
| DINO-R50 +add ‡ | 0.587 | 0.105 | 0.228 | 0.596 | 0.292 | 0.591 [0.38, 0.73] | 0.289 | 24 |
| DINO-Swin-L +add ‡ | 0.609 | 0.071 | 0.211 | 0.655 | 0.293 | 0.679 [0.50, 0.80] | 0.303 | 7 |
| MaskDINO-R50 † | 0.501 | 0.164 | 0.233 | 0.555 | 0.322 | 0.542 [0.25, 0.73] | 0.315 | 31 |
| MaskDINO-R50 +add † | 0.588 | 0.187 | 0.248 | 0.608 | 0.315 | 0.629 [0.41, 0.78] | 0.325 | 32 |
| Model | [95% CI] | FPS | ||||||
|---|---|---|---|---|---|---|---|---|
| Mask2Former-R50 | 0.534 | 0.067 | 0.197 | 0.629 | 0.276 | 0.616 [0.33, 0.80] | 0.270 | 32 |
| Mask2Former-R50 +add | 0.493 | 0.069 | 0.182 | 0.594 | 0.265 | 0.560 [0.35, 0.69] | 0.250 | 33 |
| Mask2Former-Swin-L | 0.563 | 0.072 | 0.206 | 0.630 | 0.283 | 0.630 [0.37, 0.79] | 0.283 | 23 |
| MaskDINO-R50 | 0.510 | 0.104 | 0.210 | 0.557 | 0.266 | 0.544 [0.26, 0.73] | 0.259 | 31 |
| MaskDINO-R50 +add | 0.565 | 0.133 | 0.224 | 0.593 | 0.275 | 0.601 [0.37, 0.75] | 0.279 | 32 |
| MaskDINO-Swin-L | 0.585 | 0.113 | 0.223 | 0.639 | 0.299 | 0.619 [0.36, 0.79] | 0.290 | 17 |
| Model | View | [95% CI] | ||
|---|---|---|---|---|
| Per-frame image baselines | ||||
| Mask2Former-R50 (image) | – | 0.616 [0.33, 0.80] | 0.534 | 0.197 |
| MaskDINO-R50 (image) | – | 0.544 [0.26, 0.73] | 0.510 | 0.210 |
| Mask2Former-Swin-L (image) | – | 0.630 [0.37, 0.79] | 0.563 | 0.206 |
| MaskDINO-Swin-L (image) | – | 0.619 [0.36, 0.79] | 0.585 | 0.223 |
| RF-DETR-Seg-n (image) | – | 0.656 [0.45, 0.81] | 0.651 | 0.258 |
| Configuration | [95% CI] | ||
|---|---|---|---|
| CAVIS-Swin-L, trained on RipBench | 0.613 [0.38, 0.78] | 0.601 | 0.247 |
| CAVIS-Swin-L, trained on RipVIS | 0.611 [0.35, 0.79] | 0.598 | 0.243 |
| CAVIS-R50 online, curated dense-nearest | 0.595 [0.31, 0.77] | 0.556 | 0.223 |
| CAVIS-R50 online, curated | 0.551 [0.28, 0.72] | 0.529 | 0.214 |
| CAVIS-R50 online, dense-interp | 0.566 [0.29, 0.75] | 0.530 | 0.224 |
| CAVIS-R50 online, dense-interp, round-2 ids (seed 43) | 0.573 [0.29, 0.76] | 0.537 | 0.207 |
Classification
Classification maps observations to discrete labels, including familiar, unknown, hierarchical, or multiple categories.
See all 451 papers