YOLO-v11n: Nano Real-Time Object Detector

• YOLO-v11n is an ultra-efficient object detector designed for resource-constrained, real-time applications.
• It introduces innovative modules like C3k2, SPPF, and C2PSA to enhance accuracy and speed with sub-5M parameters and under 10 GFLOPs.
• Empirical evaluations show superior performance on benchmarks, achieving high mAP scores and minimal inference latency across domains.

YOLO-v11n (YOLO version 11, nano variant) is an ultra-efficient object detector engineered for real-time applications under stringent resource and latency constraints. Characterized by a parameter count generally below 5 million and a sub-5 GFLOP computational profile, YOLO-v11n extends the established YOLO backbone–neck–head paradigm with new architectural modules, notably the C3k2 bottleneck, SPPF (Spatial Pyramid Pooling Fast), and C2PSA (Convolutional block with Parallel Spatial Attention). These innovations enable v11n to achieve superior accuracy and speed compared to all prior "nano" YOLO models, positioning it as the default choice for embedded, mobile, and battery-constrained scenarios (Kotthapalli et al., 4 Aug 2025, Behzadi et al., 14 Jul 2025, Jegham et al., 2024, Hidayatullah et al., 23 Jan 2025, Khanam et al., 2024, Jiang et al., 20 Feb 2025).

1. Architectural Innovations

YOLO-v11n’s principal advances are located in three core modules:

• C3k2 Block: This bottleneck splits the input tensor along the channel dimension, processes one partition via two small (2×2) convolutions with SiLU activations and BatchNorm, then concatenates the result with an unmodified bypass stream and fuses via a 1×1 convolution. This design increases local receptive field while reducing redundancy relative to previous C2f blocks. The mathematical structure is:

$$X = [X_a; X_b], \quad Y_1 = \mathrm{Conv}_{1\times1}(X_b), \quad Y_2 = \mathrm{Conv}_{2\times2}(\mathrm{BN}(\mathrm{Conv}_{1\times1}(X_b)))$$

$$\mathrm{C3k2}(X) = \mathrm{Conv}_{1\times1}([X_a; Y_1; Y_2])$$

• SPPF Layer: A channel-reducing 1×1 convolution is followed by three serial $5\times5$ max-poolings (stride 1, "same" padding), and the outputs across scales are concatenated. A final 1×1 conv fuses the pooled context:

$$\mathrm{SPPF}(X) = \mathrm{Conv}_{1\times1}([X, \mathrm{MP}_1(X), \mathrm{MP}_2(X), \mathrm{MP}_3(X)])$$

with $\mathrm{MP}_k$ denoting $k$-times repeated maxpool.

• C2PSA Module: Two parallel $3\times3$ convolutions compute spatial-attention masks which, after sigmoid activation and averaging, modulate the feature map channel-wise. An output 1×1 conv reintegrates attended features, yielding:

$$A = \frac{\sigma(\mathrm{Conv}^1_{3\times3}(F)) + \sigma(\mathrm{Conv}^2_{3\times3}(F))}{2}, \quad F' = A \odot F, \quad \mathrm{C2PSA}(F) = F + \mathrm{Conv}_{1\times1}(F')$$

The network adopts a CSP (Cross Stage Partial) layout with three main downsampling stages and a PANet-style neck, maintaining high spatial fidelity in the nano regime (Khanam et al., 2024, Hidayatullah et al., 23 Jan 2025, Jegham et al., 2024).

2. Model Complexity and Scaling

YOLO-v11n hyper-parameters are set for minimal depth and width:

• Depth Multiple: $d=0.33$
• Width Multiple: $w=0.25$
• Input Resolution: $$\mathrm{C3k2}(X) = \mathrm{Conv}_{1\times1}([X_a; Y_1; Y_2])$$0
• Maximum Channel Width: $$\mathrm{C3k2}(X) = \mathrm{Conv}_{1\times1}([X_a; Y_1; Y_2])$$1

The resulting layer breakdown for YOLO-v11n is as follows:

Stage	Output Res.	Channels	Module Type
Input	640×640	3→16	Conv1×1
Backbone·1	320×320	16→32	C3k2
Backbone·2	160×160	32→64	C3k2
Backbone·3	80×80	64→128	C3k2, SPPF, C2PSA
Neck (PAN)	80–20×80–20	128/64/32	Up/Down+C3k2
Head	80/40/20	64/128/256	CBS+Detect

Parameter and computation summary for YOLO-v11n:

Model	Params (M)	FLOPs (G)	Size (MB)
v11n	2.3–5.0	4.5–8.2	5–12
v11s	7.5–9.8	12–20	20–38

Slight discrepancies in parameter and FLOP counts are dataset- and implementation-dependent; the architecture always targets sub-5M params and sub-10 GFLOPs (Kotthapalli et al., 4 Aug 2025, Hidayatullah et al., 23 Jan 2025, Jegham et al., 2024).

3. Training Regimen, Losses, and Optimization

Training commonly employs the following configuration:

• Optimizer: SGD with momentum 0.937, weight decay $$\mathrm{C3k2}(X) = \mathrm{Conv}_{1\times1}([X_a; Y_1; Y_2])$$2
• Learning Rate Schedule: Linear warm-up to 0.01 over 3 epochs; cosine decay thereafter
• Epochs: 80–300, dataset-dependent
• Batch Size: 16–64 (according to hardware)
• Augmentations: Mosaic, MixUp, Random HSV, flip, scaling, Cutout/Erasing (per scenario)
• Loss Function (standard for v11n; λ's typically $$\mathrm{C3k2}(X) = \mathrm{Conv}_{1\times1}([X_a; Y_1; Y_2])$$3):

$$\mathrm{C3k2}(X) = \mathrm{Conv}_{1\times1}([X_a; Y_1; Y_2])$$4

Some implementations additionally employ Distribution Focal Loss for bounding box refinement (Kotthapalli et al., 4 Aug 2025, Behzadi et al., 14 Jul 2025, Jiang et al., 20 Feb 2025).

4. Empirical Performance and Multi-Domain Results

Performance metrics for YOLO-v11n span large-scale benchmarks and specialized domains:

• COCO 2017 (640×640):
  • $$\mathrm{C3k2}(X) = \mathrm{Conv}_{1\times1}([X_a; Y_1; Y_2])$$5: 39.5%
  • $$\mathrm{C3k2}(X) = \mathrm{Conv}_{1\times1}([X_a; Y_1; Y_2])$$6: 22.3%
  • FPS (V100): 650 (1.5 ms)
  • CPU latency: 12 ms (Core i7-1185G7)
  • Model size: 5–12 MB (FP16/FP32) (Kotthapalli et al., 4 Aug 2025, Khanam et al., 2024)
• Multi-domain ODverse33 Benchmark (across 33 datasets):
  • Mean $$\mathrm{C3k2}(X) = \mathrm{Conv}_{1\times1}([X_a; Y_1; Y_2])$$7: 0.8072
  • Mean $$\mathrm{C3k2}(X) = \mathrm{Conv}_{1\times1}([X_a; Y_1; Y_2])$$8: 0.5983
  • $$\mathrm{C3k2}(X) = \mathrm{Conv}_{1\times1}([X_a; Y_1; Y_2])$$9: 0.3794
  • Throughput: 140 FPS (A100, FP16)
  • Medical: $5\times5$0 up to 80%
  • Retail, Security, Wildlife, Agricultural: $5\times5$1 0.43–0.80 (Jiang et al., 20 Feb 2025)
• Task-specific (Colorectal Polyp Detection):
  • Precision: 95.8%
  • Recall: 91.9%
  • $5\times5$2: 96.5%
  • $5\times5$3: 77.8%
  • Inference: ~80 FPS (RTX 3050 Ti), total latency ≈11 ms/image (Behzadi et al., 14 Jul 2025)

YOLO-v11n’s C2PSA module enhances small-object and occlusion robustness, explaining its gains on aerial, medical, and microscopic benchmarks (Jiang et al., 20 Feb 2025, Behzadi et al., 14 Jul 2025).

5. Comparative Analysis with Competing Nano Models

YOLO-v11n is consistently the top-performing nano model in the YOLO family across diverse tasks. Direct comparisons:

Model	Params (M)	$5\times5$4	$5\times5$5	Latency (V100 ms)
YOLO-v5n	1.8–1.9	79.9%	0.590	1.2–1.3
YOLO-v8n	3.2–3.7	37.4%	0.588*	1.7–1.8
YOLO-v9n	3.0	80.5%	0.595*	1.7
YOLO-v10n	2.5–4.0	78.7%	0.583	1.4–2.4
YOLO-v11n	2.3–5.0	80.7%	0.598	1.5–2.2

(*Average mAP@0.5:0.95 across multi-domain tasks.) YOLO-v11n’s memory and compute efficiency comes with a negligible drop in accuracy compared to larger ‘small’ or ‘medium’ variants, but offers a 2–4× reduction in latency and model size (Kotthapalli et al., 4 Aug 2025, Jiang et al., 20 Feb 2025, Jegham et al., 2024, Khanam et al., 2024).

6. Application Scenarios and Practical Deployment

Key applications for YOLO-v11n are those that require real-time detection on constrained hardware:

• Low power embedded AI (Jetson Nano/TX2, Raspberry Pi 4, ARM SoCs)
• UAVs/drones (SLAM, obstacle avoidance, live mapping)
• Mobile AR (headsets, phones—real-time overlay at ≥30 FPS)
• Smart city, edge camera analytics (≤12 MB footprint, <2 ms latency)
• Medical imaging (e.g., live polyp detection, ~80 FPS with <20 ms end-to-end latency)

Deployment guidelines include INT8 quantization (TensorRT), channel pruning (C3k2 blocks, reducing 20–30% channels), and knowledge distillation from heavier models to recover accuracy at the ultra-tiny scale. Empirical results confirm sub-1% mAP drops with quantization, and up to 20% latency improvement via structured pruning (Khanam et al., 2024).

7. Design Insights and Recommendations

YOLO-v11n’s improvements over prior nano models are principally attributed to:

• CSP-based C3k2 blocks: Shorter convolutions but deeper feature reuse, ~20% fewer parameters than C2f (Jegham et al., 2024)
• PSA-based attention (C2PSA): ~10% higher small-object recall than non-attentive counterparts (Behzadi et al., 14 Jul 2025)
• SPPF: Low-latency spatial context aggregation without costly large kernels (Khanam et al., 2024)
• PANet neck and decoupled heads: Robust multi-scale feature fusion, anchor-free regression for dense scenarios

Deployment best-practices converge on keeping to default 640×640 input, using FP16/INT8 inference, and leveraging nano augmentation recipes. For domains driven by tiny objects or severe occlusion, tuning C2PSA bottleneck ratios or extending the head with an additional (P2) scale is advised (Jiang et al., 20 Feb 2025, Jegham et al., 2024).

---

YOLO-v11n defines the current state of the art in nano-scale real-time object detection, combining lightweight architectural innovations with rigorous empirical validation over a broad range of visual domains (Jiang et al., 20 Feb 2025, Kotthapalli et al., 4 Aug 2025, Khanam et al., 2024, Jegham et al., 2024, Hidayatullah et al., 23 Jan 2025, Behzadi et al., 14 Jul 2025).