YOLOv10: Optimized Real-Time Detector

• YOLOv10 is a state-of-the-art object detection framework that integrates a CSPNet backbone, PAN neck, and spatial attention modules for efficient, real-time performance.
• Its dual-label assignment strategy enables NMS-free inference by pairing ground-truth with the highest confidence prediction, reducing latency by up to 65%.
• The model scales from ultra-light nano to high-accuracy X variants, supporting diverse applications in marine, medical, and aerial domains.

YOLOv10 is a state-of-the-art object detection framework designed to optimize both detection accuracy and inference speed, achieving real-time, end-to-end deployment across diverse, domain-specific use cases. Building upon the YOLO family, YOLOv10 introduces architectural refinements—including a CSPNet-based backbone, PAN neck, spatial attention modules, and a novel dual-label assignment strategy—that collectively improve detection performance and efficiency while eliminating non-maximum suppression (NMS) during inference. This model family scales from ultra-light "nano" variants for embedded devices to "X" variants for maximal accuracy, and is demonstrated across applications ranging from marine monitoring and agricultural analysis to medical imaging and embedded real-time systems (Wang et al., 2024, Wuntu et al., 22 Sep 2025, Ahmed et al., 2024, Hussain, 2024).

1. Architecture and Main Components

YOLOv10 retains the canonical single-stage "input→backbone→neck→head→output" pipeline, structurally organized as follows:

• Backbone (CSPNet-based): Utilizes Cross-Stage Partial (CSP) connections, splitting feature maps and processing only one part through deep convolutional stacks before recombination. YOLOv10 replaces standard bottlenecks with efficiency-focused blocks:
  • Compact Inverted Blocks (CIB): 1×1 pointwise expansion, depthwise 3×3 (or large-kernel 7×7 in deeper layers), then 1×1 pointwise projection.
  • Spatial-Channel Decoupled Downsampling: Separates 3×3 depthwise spatial downsampling and 1×1 pointwise channel projection.
  • Rank-Guided Block Design: Intrinsic rank measurement informs selective pruning or narrowing of layers.
• Neck (Path Aggregation Network): PANet fuses multiscale features via decoupled spatial–channel operations, cross-stage aggregation, and partial self-attention (PSA) modules at the finest scales for small-object sensitivity.
• Heads:
  • Dual-Head Structure: Both one-to-many and one-to-one assignment heads during training drive dense coverage (recall) and precise localization (precision), respectively. Inference exclusively utilizes the one-to-one head, ensuring a single prediction per ground-truth box.
  • Decoupled Design: Distinct branches for bounding-box regression, classification, and objectness, typically employing lightweight or depthwise-separable convolutions.
• Anchor-Free Predictions: Each feature map cell predicts the absolute shape and location of objects, eliminating the need for predefined anchor boxes.
• Attention Modules: PSA blocks or equivalent attention modules may be inserted for improved spatial awareness and global context at minimal overhead (Wuntu et al., 22 Sep 2025, Wang et al., 2024, Hussain, 2024, Tian et al., 2024).

2. Dual-Label Assignment and NMS-Free Inference

A central innovation of YOLOv10 is the consistent dual assignment strategy for NMS-free training and inference:

• One-to-Many Assignment: Each ground-truth box is associated with multiple positive predictions (whose anchor/grid center lies in the box), enhancing gradient diversity and recall.
• One-to-One Assignment: Each ground-truth box is paired with the prediction that maximizes a unified match score:

$$m(\alpha, \beta) = s \cdot p^\alpha \cdot \text{IoU}(b, b^*)^\beta$$

where $s$ is a spatial prior, $p$ is predicted class score, and $\text{IoU}(b, b^*)$ is intersection-over-union with hyperparameters $\alpha, \beta$ set per head (Wang et al., 2024, Ahmed et al., 2024).

• Inference: Only the one-to-one head is retained. Since it guarantees a unique, highest-confidence prediction for each object, costly post-processing via NMS is not required, reducing end-to-end latency by 60–65% (Wang et al., 2024, Hussain, 2024).

3. Efficiency-Accuracy Optimizations and Model Scaling

YOLOv10 is systematically optimized at both micro- and macro-architecture levels:

• Spatial–Channel Decoupling: Replaces monolithic stride-2 convolutions with separate spatial and channel operations, reducing multiply-accumulate operations (MACs) by ~30% compared to YOLOv9 backbones (Alif et al., 2024, Wang et al., 2024).
• Rank-Guided Pruning: Analytical intrinsic rank evaluation allows removal or narrowing of low-information blocks, mainly in the backbone.
• Lightweight Decoupled Heads: Heads use depthwise-separable or simplified 1×1 convolutions.
• Support for Compound Scaling: Five main variants—Nano (N), Small (S), Medium (M), Large (L), and eXtra-Large (X)—are defined, trading off parameter count, capacity, and latency. Parameter and FLOP counts for common variants (from (Ahmed et al., 2024, Wang et al., 2024)):

Variant	Parameters (M)	FLOPs (G)	COCO mAP@50:95 (%)	Latency (ms, T4/TRT)
N	2.7	6.7–8.2	39.1–39.5	~1.8
S	7.2–8.0	21.6–24.5	46.3–46.8	~2.4
M	15.4–16.5	59.1–63.5	51.3–51.9	~4.6
L	24.4–25.7	120–126	53.4–46.6–58.8*	~7–10
X	29.5–31.6	160–169	54.4–48.2–57.8*	~10–12

*L/X class-specific detection precision reported for medical tasks (Ahmed et al., 2024, Wang et al., 2024, Hussain, 2024).

• Edge Deployment Considerations: INT8-quantized nano variants fit within a ~5 MB memory footprint, achieving real-time (20–30 FPS) inference on embedded hardware (e.g., Jetson Xavier NX) (Hussain, 2024, Wuntu et al., 22 Sep 2025).

4. Training Pipeline, Hyperparameters, and Losses

• Standard Training Procedure:
  • Input size: typically 640×640.
  • Augmentation: mosaic, mixup, random affine, color jitter, label smoothing, and application-specific additions (e.g., random erasing, perspective warping).
  • Optimizers: SGD or AdamW; momentum ≈ 0.9–0.937; weight decay = 5e−4.
  • Learning Rate Schedules: cosine annealing or flat schedule; warmup epochs typical (Ahmed et al., 2024, Wang et al., 2024, Wuntu et al., 22 Sep 2025).
• Loss Function Structure:

$$\mathcal{L} = \lambda_{\text{box}}\mathcal{L}_{\text{box}} + \lambda_{\text{cls}}\mathcal{L}_{\text{cls}} + \lambda_{\text{dfl}}\mathcal{L}_{\text{dfl}} + \lambda_{\text{obj}}\mathcal{L}_{\text{obj}}$$

• Bounding Box Regression: Complete IoU/Distance IoU/Generalized IoU.
• Classification: Task-aligned focal loss, BCE.
• Distribution Focal Loss (DFL): For accurate continuous localization.
• Objectness: BCE (Wang et al., 2024, Ahmed et al., 2024).
  • Task-Specific Weights and Variations: In training for medical, drone, or micro-object tasks, additional focal scaling or custom regression losses (e.g., Wise-IoU v3) may be integrated for robustness to annotation noise or label imbalance (Farooqui et al., 13 Feb 2026, Ahmed et al., 2024).

5. Quantitative Performance and Comparative Benchmarks

YOLOv10 establishes state-of-the-art balance between accuracy and speed:

• COCO Benchmarks: YOLOv10-S achieves 46.3% COCO [email protected]:0.95, 2.49 ms latency, outperforming analogously sized YOLOv8/YOLOv9 in both throughput and detection precision (Wang et al., 2024, Hussain, 2024, Alif et al., 2024).
• Specialized Domains:
  • Marine Fish Detection: YOLOv10-nano (2.7M params) reaches mAP@50 = 0.966, mAP@50:95 = 0.606 on DeepFish with 29.3 FPS on CPU (Wuntu et al., 22 Sep 2025).
  • Medical Imaging: YOLOv10-M achieves mAP@50:95 = 51.9% (GRAZPEDWRI-DX wrist X-rays), an 8.6 pp improvement over YOLOv9-E (Ahmed et al., 2024).
  • Drone Object Detection: LAF-YOLOv10 attains 35.1% [email protected] on VisDrone-DET2019 with 2.3M parameters (Farooqui et al., 13 Feb 2026).
  • Retail Self-Checkout: Models integrating YOLOv8-style heads on the YOLOv10 backbone report [email protected] = 0.871 with under 10 GFLOPs (Tan et al., 2024).
  • Ecology and Bio-Imaging: Achieves [email protected] ≈ 0.96 on herbarium plant regions; 0.976 for multi-species bird monitoring (Sklab et al., 22 Jul 2025, Chalmers et al., 2024); [email protected] ≈ 0.990 for blood-cell detection (Choudhary et al., 2024).

Application	Variant	mAP@50	mAP@50:95	Parameters	FPS/CPU	Reference
DeepFish	YOLOv10-n	0.966	0.606	2.7 M	29.29	(Wuntu et al., 22 Sep 2025)
GRAZPEDWRI-DX	YOLOv10-M	—	0.519	16.5 M	—	(Ahmed et al., 2024)
VisDrone-DET2019	LAF-YOLOv10	0.351	0.207	2.3 M	24.3 (FP16)	(Farooqui et al., 13 Feb 2026)
Plant region	YOLOv10	0.959	—	—	—	(Sklab et al., 22 Jul 2025)

6. Domain-Specific Variations and Extensions

Numerous studies adapt the YOLOv10 core framework to address domain-specific challenges:

• Marine/Aquatic Imaging: Add-ons include module replacements for enhanced small-object accuracy (e.g., FasterNet backbone, compact detection heads) leading to significant AP gains in detecting dead fish on wide water surfaces (Tian et al., 2024).
• Aerial/Drone Imagery: Integration of partial convolution blocks, attention-guided fusion, and high-resolution auxiliary heads (e.g., P2 at 160×160 pixels), as well as robust loss reweighting (Wise-IoU v3), address detection of <8×8 px targets (Farooqui et al., 13 Feb 2026).
• Medical/Surgical Video: Architecture is minimally modified, with classification heads adapted for multi-task outputs (e.g., hand laterality), extensive augmentation for robustness, and dropout for generalization (Sun et al., 21 Feb 2026).
• Retail and Embedded Systems: Detection head may be reverted to YOLOv8-style point-based architectures with standard NMS for optimal product detection and system integration (Tan et al., 2024).

7. Limitations, Discussion, and Future Research Directions

• Model Scaling: Performance saturates or degrades beyond medium/large variants due to redundancy, particularly for small-structure detection (Ahmed et al., 2024).
• Training Convergence: The one-to-one head alone converges slowly—dual assignment is essential for both recall and precision (Wang et al., 2024).
• NMS-Free Trade-offs: While end-to-end latency is significantly reduced, the smallest models (nano, small) may still incur a 0.5–1.0 AP deficit versus standard NMS-trained models (Wang et al., 2024).
• Generalization to Other Architectures: The dual-label assignment and other YOLOv10-specific optimizations have yet to be systematically evaluated across alternative backbone or attention designs (Ahmed et al., 2024).
• Emerging Research Directions: Areas identified for further investigation include dynamic assignment metrics, advanced self-distillation, adaptive attention modules, multi-modal input fusion (e.g., RGB+NIR+thermal), online adaptation for evolving domains, and interpretability via attention mapping (Alif et al., 2024, Wang et al., 2024).

• (Wang et al., 2024) YOLOv10: Real-Time End-to-End Object Detection
• (Wuntu et al., 22 Sep 2025) Real-Time Fish Detection in Indonesian Marine Ecosystems Using Lightweight YOLOv10-nano Architecture
• (Ahmed et al., 2024) Pediatric Wrist Fracture Detection in X-rays via YOLOv10 Algorithm and Dual Label Assignment System
• (Hussain, 2024) YOLOv5, YOLOv8 and YOLOv10: The Go-To Detectors for Real-time Vision
• (Tian et al., 2024) A method for detecting dead fish on large water surfaces based on improved YOLOv10
• (Sklab et al., 22 Jul 2025) PlantSAM: An Object Detection-Driven Segmentation Pipeline for Herbarium Specimens
• (Tan et al., 2024) Enhanced Self-Checkout System for Retail Based on Improved YOLOv10
• (Farooqui et al., 13 Feb 2026) LAF-YOLOv10 with Partial Convolution Backbone, Attention-Guided Feature Pyramid, Auxiliary P2 Head, and Wise-IoU Loss for Small Object Detection in Drone Aerial Imagery
• (Chalmers et al., 2024) AI-Driven Real-Time Monitoring of Ground-Nesting Birds: A Case Study on Curlew Detection Using YOLOv10
• (Choudhary et al., 2024) Transforming Blood Cell Detection and Classification with Advanced Deep Learning Models: A Comparative Study
• (Alif et al., 2024) YOLOv1 to YOLOv10: A comprehensive review of YOLO variants and their application in the agricultural domain