Research map / Detection and segmentation / CNN features
AMSP-UOD: When Vortex Convolution and Stochastic Perturbation Meet Underwater Object Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AMSP-UOD, a single-stage underwater object detector with Amplitude-Modulated Shuffling Perturbation Vortex Convolution (AMSP-VConv), a FAD-CSP neck, and aspect-ratio-similarity NMS, evaluated on URPC and RUOD.
Problem
Underwater object detection suffers from noise, light absorption and scattering, small clustered objects, and class imbalance, which degrade CNN detector accuracy and robustness.
Approach
Introduces AMSP-VConv, which randomly perturbs channel groupings (amplitude modulation and shuffling) followed by shared-weight group convolutions to disrupt noise patterns while reducing parameters; adds a global feature-aware strip-pooling attention in a CSP block (FAD-CSP) and an NMS variant that skips aspect-ratio-dissimilar boxes during suppression.
Key results
- On URPC, Ours (AMSP-VConv + NMS-Similar) reaches 40.1 AP and 78.5 AP50, versus 38.6 AP / 66.2 AP50 for YOLOv5s; Ours (Standard) reaches 45.0 AP on URPC.
- On RUOD, the full model reaches 65.2 AP and 86.1 AP50, above the best listed baseline TOOD at 55.3 AP and 83.1 AP50.
- Under Gaussian noise, the method's mAP stays stable at noise level 4 while YOLOv5s mAP@0.5 drops 16.3%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add AMSP-VConv and NMS-Similar | AMSP-UOD · URPC | AP50: 73.4 -> 78.5 | Helped |
| Add AMSP-VConv and NMS-Similar | AMSP-UOD · RUOD | AP: 61.4 -> 65.2 | Helped |
| Add AMSP-VConv | AMSP-UOD · URPC | AP: 45.0 -> 36.6 | Hurt |
| Add AMSP strategy to VConv (ablation) | VConv · URPC | mAP@[0.5:0.05:0.95]: 0.397 -> 0.398 | Helped |
| Add shared weights to VConv (ablation) | VConv · URPC | mAP@[0.5:0.05:0.95]: 0.386 -> 0.398 | Helped |
| Use NMS-Similar instead of NMS (ablation) | underwater object detection post-processing · Underwater detection | mAP@[0.5:0.05:0.95]: 0.366 -> 0.401 | Helped |
| Use NMS-Similar instead of Soft-NMS (ablation) | underwater object detection post-processing · Underwater detection | inference time: 337.3 -> 46.90 | Helped |
| Remove GFA from FAD-CSP (ablation) | FAD-CSP · Underwater detection | mAP@0.5: 0.675 -> 0.748 | Hurt |
| Use pooling groups instead of individual pooling layers (ablation) | FAD-CSP · Underwater detection | mAP@0.5: 0.640 -> 0.748 | Helped |
| Add AMSP strategy to GFA (ablation) | FAD-CSP · Underwater detection | mAP@0.5: 0.679 -> 0.748 | Helped |
| Use RepBottleneck instead of Bottleneck (ablation) | FAD-CSP · Underwater detection | mAP@0.5: 0.681 -> 0.748 | Helped |
| Use AMSP-VConv rather than standard convolutions (ablation) | AMSP-UOD · URPC with Gaussian noise | mAP@0.5: null | Helped |
Surprising findings
- The Standard variant has the highest URPC AP (45.0) yet lower AP50 than the AMSP-VConv variants, and the AMSP-VConv version is slightly lower than Standard on balanced metrics, which the authors acknowledge.
Limitations
- Performance is inconsistent across variants; AMSP-VConv version trails the Standard variant on several metrics.
- Ablation tables report inconsistent or unclear metrics and only partial component breakdowns.
- Evaluated only on URPC and RUOD underwater datasets.
Experimental setup
Datasets: URPC RUOD
Benchmarks: URPC RUOD
Metrics: AP AP50 precision recall mAP@0.5 inference time
Baselines: YOLOv3 YOLOv5s YOLOv6s YOLOv7-tiny Faster R-CNN Cascade R-CNN RetinaNet FCOS ATSS TOOD PAA Soft-NMS
Builds on: Soft-NMS CSP YOLO
Closest papers by meaning
- WaterDiffusion: Learning a Prior-involved Unrolling Diffusion for Joint Underwater Saliency Detection and Visual Restoration (AAAI 2025 · similarity 0.84)
- UVEB: A Large-scale Benchmark and Baseline Towards Real-World Underwater Video Enhancement (CVPR 2024 · similarity 0.83)
- WebUOT-1M: Advancing Deep Underwater Object Tracking with A Million-Scale Benchmark (NeurIPS 2024 · similarity 0.83)
- Leveraging Perturbation Robustness to Enhance Out-of-Distribution Detection (CVPR 2025 · similarity 0.83)
- HSOD-BIT-V2: A Challenging Benchmark for Hyperspectral Salient Object Detection (AAAI 2025 · similarity 0.81)
- AdaptiveISP: Learning an Adaptive Image Signal Processor for Object Detection (NeurIPS 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.