Abstract
Underwater computer vision faces unique challenges such as light attenuation, turbidity, and class imbalance. This study provides a controlled comparison of the latest YOLO iterations (v8 through v11) specifically for marine environments. While detection accuracy has largely plateaued since YOLOv9, architectural innovations in YOLOv10 offer superior speed-accuracy trade-offs essential for real-time deployment on autonomous underwater vehicles (AUVs).
Introduction
Computer vision in the marine domain is notoriously difficult. Unlike the clear, well-lit conditions of terrestrial benchmarks like COCO or PASCAL-VOC, underwater imagery suffers from significant optical degradation. AUVs require models that are not only robust against these "blurry" conditions but also efficient enough to run on low-power embedded hardware. This paper benchmarks the newest YOLO releases (v9, v10, and v11) against the established YOLOv8 to determine their suitability for the unique challenges of the ocean.
Methodology
The evaluation focused on the "small" (-s) variants of YOLOv8, v9, v10, and v11. All models were trained from scratch using identical hyperparameters: 100 epochs, 640px input resolution, and a batch size of 16 on a Tesla T4 GPU.
A key aspect of the methodology was testing data efficiency. The researchers created four training regimes—using 25%, 50%, 75%, and 100% of the available training images—to observe how each architecture scales with limited data. Furthermore, Grad-CAM (Gradient-weighted Class Activation Mapping) was employed to visualize the regions of interest for the models, providing explainability beyond raw performance metrics.
Datasets
Two open-access marine datasets were utilized:
- Coral Disease Dataset: 4,480 images across 18 classes, testing the ability to distinguish subtle textures in sessile organisms.
- Fish Species Dataset: 7,500 images across 20 classes, simulating traditional detection with moving targets and varying light.
Both datasets were manually verified and balanced to ensure 20 samples per class in the validation and test sets, preventing evaluation bias.
Performance Analysis
Accuracy and Data Sensitivity
One of the most striking findings is that accuracy (mAP) effectively saturates after YOLOv9. Despite the "v11" label, newer models did not show significant jumps in precision or recall for these specific marine tasks. However, the models demonstrated high stability across different training splits. Even at 25% data, performance remained respectable, although lower-data models were more prone to focusing on irrelevant background elements.
Efficiency and Real-time Deployment
While accuracy plateaued, efficiency gains were notable. YOLOv10-s emerged as the optimal choice for AUV deployment, providing the highest Frames Per Second (FPS) and lowest inference time. It achieved approximately 43.92 FPS on the Coral dataset and 43.51 FPS on the Fish dataset. Surprisingly, YOLOv11-s was slightly slower than v10 in these tests (~41-43 FPS). YOLOv8 was significantly slower, taking nearly double the inference time per image compared to v10.
Explainability via Grad-CAM
Grad-CAM analysis revealed that even high-performing models often rely on "spurious features"—background noise or water texture—rather than the actual marine organism. This "distraction factor" is a critical consideration for model reliability in diverse underwater environments.
Conclusion and Future Work
The study concludes that YOLOv10-s offers the best balance of speed and accuracy for underwater tasks. Future work will involve expanding the benchmark to include all YOLO versions, developing better explainability metrics for grid-based detectors, and testing models in "live" underwater environments to assess temporal consistency for AUV navigation.
Summary updated