Does YOLO26 Truly Offer Advantages Over Its Predecessors for Edge Deployment? A Benchmark Study in Aquaculture

Technical Summary: Benchmarking YOLO26 for Edge Deployment in Aquaculture

Problem Statement
The adoption of Computer Vision (CV) and Artificial Intelligence (AI) in intensive aquaculture, particularly Recirculating Aquaculture Systems (RAS), offers significant potential for automated monitoring of fish health and welfare. However, deploying these systems faces unique challenges, including high stocking densities, water turbidity, optical distortions, and the rapid, unpredictable motion of aquatic animals. A critical constraint is the reliance on resource-constrained edge devices (e.g., CPU-based single-board computers like the Raspberry Pi) due to limited bandwidth for cloud transmission and the high cost/power consumption of GPU-accelerated embedded platforms.

While the You Only Look Once (YOLO) family of models is widely used for real-time object detection, the performance of the latest generation, YOLO26, remains unvalidated in domain-specific aquaculture scenarios. YOLO26 introduces architectural innovations such as Non-Maximum Suppression (NMS)-free end-to-end inference and optimizations for CPU-based edge deployment. The central question addressed by this study is whether these architectural advancements translate to tangible advantages over predecessors (YOLOv5u, YOLOv8, and YOLO11) for the specific task of detecting fish mortality in RAS environments, or if model selection should be driven by other factors such as data availability and hardware constraints.

Methodology
This study conducted a comprehensive benchmarking evaluation of twelve Ultralytics YOLO model variants (four generations: v5u, v8, 11, 26; across three scales: nano, small, medium) for the detection of fish mortality (dead vs. live).

• Dataset: A semi-commercial-scale RAS dataset (RAS-MortDB) was utilized, comprising 2,000 manually annotated images of Atlantic salmon (Salmo salar) collected over 90 days. The dataset included diverse lighting conditions (ambient and supplemental) and mortality levels. Images were augmented to create a training set of 2,800 images with 2,204 'dead' and 3,209 'live' instances.
• Training Environment: Models were trained on an NVIDIA A100 GPU cluster using standardized hyperparameters (100 epochs, 640px input, batch size 32) to ensure fair comparison.
• Evaluation Metrics:
  1. Accuracy: Mean Average Precision at IoU 0.50 (mAP50) and 0.50–0.95 (mAP50-95).
  2. Data Efficiency: Learning curve analysis across seven dataset sizes (100 to 2,800 images) to determine minimum data requirements for 90% mAP50.
  3. GPU Performance: Inference speed (FPS) and latency on A100 GPUs, distinguishing between raw forward-pass time, post-processing time (NMS overhead), and end-to-end latency.
  4. Edge Deployment: Inference benchmarking on a Raspberry Pi 5 (CPU-only, 8GB RAM) using ONNX Runtime. Metrics included average latency, FPS, thermal throttling status, and model size (ONNX footprint).

Key Results

• Detection Accuracy: All 12 models achieved comparable detection accuracy on the full dataset, with mAP50 values ranging narrowly from 93.77% (YOLOv8m) to 94.81% (YOLO11m). The difference of only 1.04 percentage points indicates that architectural generation has minimal influence on final mortality detection accuracy when sufficient training data is available.
• Data Efficiency: YOLOv8 demonstrated superior training efficiency, achieving ≥90% mAP50 with only 400 training images. In contrast, YOLO26 (nano and small variants) required 1,000 images to reach the same threshold, making it approximately 2.5 times less data-efficient. This is attributed to YOLO26's one-to-one label assignment strategy, which yields sparser supervisory signals compared to the many-to-one strategies of earlier models.
• GPU Inference: On high-performance GPUs, YOLOv8 consistently achieved the highest inference speeds (FPS) and lowest end-to-end latency. Despite YOLO26's NMS-free design reducing post-processing overhead significantly, it incurred higher residual latency (preprocessing and tensor allocation), resulting in lower overall FPS compared to YOLOv8 and YOLOv5u.
• Edge (CPU) Inference: Performance rankings reversed on the resource-constrained Raspberry Pi 5:
  • Nano/Small Tiers: YOLO26n achieved the highest inference speed (7.51 FPS), outperforming YOLOv8n (6.51 FPS) and YOLO11n (6.71 FPS). This validates YOLO26's NMS-free design for CPU efficiency.
  • Medium Tier: Surprisingly, YOLOv5mu outperformed all other medium-scale models (1.25 FPS vs. ~1.05 FPS for YOLO26m), likely due to the structural simplicity of its anchor-based architecture being more compatible with ARM CPU instruction sets than the complex attention mechanisms of newer models.
• Model Size: YOLO26n produced the smallest ONNX export (9.8 MB), offering advantages for storage-constrained edge devices.

Key Contributions

1. Domain-Specific Validation: This study provides the first peer-reviewed benchmark of YOLO26 in an aquaculture context, moving beyond general-purpose benchmarks (e.g., COCO) to validate performance on real-world fish mortality data.
2. Hardware-Agnostic Insights: The research highlights a critical divergence between GPU and CPU performance rankings, demonstrating that GPU benchmarks are insufficient proxies for edge deployment suitability.
3. Data Efficiency Analysis: The study quantifies the trade-off between architectural novelty and data requirements, revealing that newer architectures (YOLO26) may demand significantly larger annotated datasets to converge to the same performance levels as established models (YOLOv8).
4. Open-Source Resources: The authors have committed to releasing the RAS-MortDB dataset and all trained model weights (PyTorch and ONNX) to facilitate reproducibility and further research.

Significance and Claims
The paper concludes that architectural novelty alone is an insufficient criterion for model selection in aquaculture applications. While YOLO26 offers distinct advantages for CPU-based edge deployment (specifically in nano/small tiers) due to its NMS-free inference, these benefits come with trade-offs in data efficiency and training time.

The authors assert that model selection should be guided by a deployment-oriented framework considering:

• Data Availability: If annotated data is scarce, YOLOv8 is preferable due to faster convergence.
• Target Hardware: For CPU-constrained edge devices, YOLO26 (nano/small) or YOLOv5 (medium) may offer superior inference speeds compared to newer generations.
• Inference Requirements: High-throughput GPU tasks may favor YOLOv8, while low-power edge tasks may benefit from YOLO26's streamlined design.

The study emphasizes that for the specific task of fish mortality detection in RAS, where events develop over minutes rather than seconds, the marginal accuracy differences between models are negligible. Therefore, practical constraints—such as the cost of annotation, hardware limitations, and energy consumption—should drive deployment decisions rather than the pursuit of the latest architectural generation. The authors caution against relying solely on GPU benchmarks and advocate for rigorous validation on target edge hardware to ensure operational viability.