SFN-YOLO: Scale-Aware Poultry Detector

• SFN-YOLO is a lightweight real-time detector designed for free-range poultry detection, effectively addressing scale variation and occlusion challenges.
• The Scale-aware Fusion Module (SFM) integrates local convolutional detail with global self-attention to refine multiscale feature representations.
• Evaluations on the M-SCOPE dataset show SFN-YOLO achieves high mAP with fewer parameters and demonstrates robust cross-domain generalization.

SFN-YOLO, abbreviated from “Scale-aware Fusion Network for YOLO,” is a lightweight real-time object detector built on YOLOv8 for free-range poultry detection. It is proposed for environments in which poultry appear at multiple scales, are frequently occluded, and are embedded in complex or dynamic backgrounds; the method addresses these conditions by combining detailed local features with broader global context through scale-aware fusion. The same work introduces the M-SCOPE dataset for varied free-range conditions and reports that SFN-YOLO achieves an mAP of 80.7% with 7.2M parameters while retaining strong generalization capability across different domains (Chen et al., 21 Sep 2025).

1. Problem setting and motivation

SFN-YOLO is framed around the practical requirements of smart poultry farming rather than generic object detection alone. The target setting is free-range poultry detection, where chickens may appear very small, medium, or large depending on distance and viewpoint, may be partially hidden by other birds or structures, and may be difficult to separate from grass, fences, shadows, equipment, or dynamic lighting. The paper further identifies domain shift across farms and scenes as a central failure mode for detectors trained in one environment and deployed in another (Chen et al., 21 Sep 2025).

The work argues that the task is not resolved by simply selecting an existing detector. High-accuracy detectors such as Faster R-CNN are described as too slow for deployment, whereas lightweight detectors remain fast but may lose precision on small or partially hidden birds. Existing poultry-focused methods are also characterized as being developed mainly for controlled laboratory, caged, or simple indoor environments rather than open free-range scenes. Consequently, the operational objectives are fourfold: real-time inference, small-object sensitivity, robustness to occlusion and clutter, and cross-scene generalization.

The paper states three main contributions. First, it introduces SFN-YOLO as an efficient scale-aware poultry detector based on YOLOv8. Second, it proposes a Scale-aware Fusion Module, or SFM, that combines local convolutional detail with global self-attention context through a bidirectional guidance strategy. Third, it presents M-SCOPE, a new dataset with 6,600 images from diverse free-range scenes, intended to benchmark detection under clutter, scale variation, and domain shift (Chen et al., 21 Sep 2025).

2. Architectural design and the Scale-aware Fusion Module

SFN-YOLO is based on YOLOv8, but the original C2f blocks in the backbone are replaced with SFM blocks. The paper places SFM at three backbone stages before the neck. This is explicitly described as a pre-fusion mechanism intended to refine representations before feature pyramid aggregation, strengthening multiscale representation through both local precision and global awareness (Chen et al., 21 Sep 2025).

The input to SFM is a feature map

$X^{i} \in \mathbb{R}^{C \times H \times W}.$

The module has three components: parallel feature extraction, bidirectional guidance, and feature fusion.

In the local convolutional branch, the model extracts fine-grained spatial detail relevant to edges, textures, small chickens, object boundaries, and partially occluded instances. This branch consists of two $3 \times 3$ convolution layers, each followed by BN and SiLU, and is written as

$X^{l} = \mathcal{C}_{3 \times 3}(X^{i}).$

In the global self-attention branch, the feature map is flattened into a token sequence

$$X^{f} \in \mathbb{R}^{N \times C}, \quad N = H \times W.$$

Queries, keys, and values are projected from $X^{f}$. The paper applies L2 normalization to $Q$ and $K$, making their dot product behave like cosine similarity, and replaces the usual fixed scaling with a learnable temperature parameter $\gamma$. The attention is defined as

$$\text{Attention}(Q, K, V) = \text{Softmax}\left(\frac{\mathcal{N}(Q)\mathcal{N}(K)^T}{\gamma}\right)V.$$

Using $\mathcal{A}$ for LayerNorm + MHSA and $3 \times 3$0 for LayerNorm + FFN, the branch is written as

$3 \times 3$1

followed by

$3 \times 3$2

after which $3 \times 3$3 is reshaped back to a 2D feature map $3 \times 3$4.

The distinguishing mechanism is bidirectional guidance between the two branches rather than simple parallel processing. Spatial guidance from local to global is defined by

$3 \times 3$5

where $3 \times 3$6 is sigmoid and $3 \times 3$7. This spatial weighting map emphasizes important pixels and regions so that the global branch focuses on poultry locations rather than irrelevant background.

Channel guidance from global to local is derived from a channel descriptor obtained by global average pooling:

$3 \times 3$8

A two-layer mapping then produces channel attention,

$3 \times 3$9

where $X^{l} = \mathcal{C}_{3 \times 3}(X^{i}).$0 is implemented with two $X^{l} = \mathcal{C}_{3 \times 3}(X^{i}).$1 convolutions and GELU, yielding $X^{l} = \mathcal{C}_{3 \times 3}(X^{i}).$2.

The paper then applies the guidance maps cross-branch and fuses the outputs with a residual connection:

$X^{l} = \mathcal{C}_{3 \times 3}(X^{i}).$3

The stated motivation is that local and global branches address complementary failure modes: the local branch preserves edges, textures, and fine boundaries for small targets and precise localization, whereas the global branch supplies semantic priors, reconstructs occluded shapes, and suppresses confusing background patterns. The bidirectional guidance is intended to reduce noisy or irrelevant channels while also focusing attention on correct spatial locations (Chen et al., 21 Sep 2025).

3. Training objective, implementation, and evaluation protocol

The paper states that SFM adopts the YOLOv8 loss function. Concretely, the composite objective combines CIoU for localization, BCE for classification, and DFL for distribution-based regression. No new detection loss is introduced; the architectural modification lies in the feature extraction and fusion mechanism rather than in an alternative optimization criterion (Chen et al., 21 Sep 2025).

The reported training configuration uses NVIDIA GeForce RTX 3090 hardware and the PyTorch framework. Images are resized to $X^{l} = \mathcal{C}_{3 \times 3}(X^{i}).$4, training runs for 300 epochs, and optimization uses SGD with momentum $X^{l} = \mathcal{C}_{3 \times 3}(X^{i}).$5, weight decay $X^{l} = \mathcal{C}_{3 \times 3}(X^{i}).$6, batch size 2, and initial learning rate $X^{l} = \mathcal{C}_{3 \times 3}(X^{i}).$7. Early stopping patience is 50, and the self-attention branch uses 8 heads.

The evaluation compares SFN-YOLO against Faster R-CNN, RetinaNet, YOLOX, YOLOv5n / YOLOv5s, YOLOv8n / YOLOv8s, YOLO11n / YOLO11s, RT-DETR-R50 / R101, and MambaYOLO-T / B. Reported metrics include mAP, AP50, AP75, APS, FPS, and Params. For out-of-domain evaluation, the paper also reports ARM, described in the table as Average Recall Metric (Chen et al., 21 Sep 2025).

A plausible implication is that the paper is designed to demonstrate not only absolute accuracy but the accuracy-efficiency-generalization trade-off that is especially salient in deployment-oriented agricultural detection.

4. M-SCOPE dataset

M-SCOPE, short for “Multiscale Cluttered Occluded Poultry Environment,” is introduced as a new dataset for free-range poultry object detection with the “chicken” class annotated. The dataset was collected from 14 scenes and includes both indoor and outdoor free-range settings. Outdoor videos were recorded from January to April 2025 at a capture resolution of $X^{l} = \mathcal{C}_{3 \times 3}(X^{i}).$8 and a frame rate of 24 fps. Annotation was performed using T-Rex Label, and the annotation format follows the VOC2011 standard (Chen et al., 21 Sep 2025).

The dataset contains 6,600 images and is split under a closed-set protocol intended to study generalization. The in-domain split uses 10 outdoor free-range scenes for training, validation, and testing, while the out-of-domain split comprises 4 novel scenes used only for testing.

Split	Images	Boxes
ID train	4000	68,850
ID val	500	8,642
ID test	500	8,932
OOD outdoor1	500	7,845
OOD outdoor2	500	4,076
OOD indoor	500	8,327
OOD snow	100	386

The dataset also reports box size distributions. In the ID train split, the distribution is S 4.20%, M 37.54%, L 58.26%. In OOD outdoor1, it is S 0.04%, M 30.04%, L 69.92%; in OOD indoor, S 6.14%, M 22.43%, L 71.43%. These figures indicate that medium and large objects dominate, but a nontrivial small-object subset remains present.

The paper presents M-SCOPE as significant because it explicitly captures multiscale, cluttered, occluded, and domain-shifted free-range poultry environments that existing datasets do not adequately represent. This suggests that the benchmark is intended to stress both representation quality and transfer robustness rather than only in-domain fitting performance (Chen et al., 21 Sep 2025).

5. Quantitative performance, ablation, and cross-domain behavior

On M-SCOPE, the paper reports the following headline results for SFN-YOLO variants and selected baselines (Chen et al., 21 Sep 2025).

Model	mAP	Params / FPS
SFN-YOLOs	80.7%	7.2M / 112.4
YOLOv8s	80.3%	11.1M / —
SFN-YOLOn	78.8%	2.0M / 116.3
YOLOv5s	80.1	—
YOLO11s	80.1	—
Faster R-CNN	63.2	41.3M / 34.4
RetinaNet	59.4	19.8M / 41.6
RT-DETR-R50	71.2	—
MambaYOLO-B	65.0	—

For SFN-YOLOs, the reported metrics are mAP 80.7%, AP50 96.8%, AP75 90.0%, APS 75.8%, FPS 112.4, and Params 7.2M. Relative to the YOLOv8s baseline, the reported gain is +0.4 mAP with 35.1% fewer parameters, from 11.1M to 7.2M, while retaining real-time speed. SFN-YOLOn records mAP 78.8%, AP50 96.7%, AP75 88.9%, APS 74.9%, FPS 116.3, and Params 2.0M, and is highlighted as the most parameter-efficient model in the table.

The ablation study uses YOLOv8n as the baseline. The baseline metrics are mAP 78.08, AP75 87.38, and APS 72.38. Adding only the global branch reduces mAP to 77.67. Combining global and local branches increases mAP to 78.19. Adding channel guidance yields 78.10, while adding spatial guidance yields 78.13. The full model with GB + LB + C + S reaches mAP 78.79, AP50 96.66, AP75 88.86, APS 74.87, APM 73.04, ARS 82.40, and ARM 80.09. The paper interprets these results as showing that the global and local branches are complementary, that channel and spatial guidance alone may introduce bias, and that both guidance mechanisms together give the best result (Chen et al., 21 Sep 2025).

For cross-domain evaluation, the ID-trained model is tested on Outdoor1 OOD data. SFN-YOLOn achieves mAP 73.2%, AP50 91.2%, AP75 80.5%, and ARM 67.4. The cited baselines on the same OOD split are YOLOv5n at 72.8 mAP, YOLOv8n at 72.6 mAP, YOLO11n at 72.4 mAP, RT-DETR-R50 at 69.1 mAP, and MambaYOLO-T at 59.0 mAP. The paper therefore presents SFN-YOLO as generalizing slightly better than YOLO baselines and substantially better than the listed transformer or state-space baselines on this unseen outdoor domain (Chen et al., 21 Sep 2025).

The qualitative analysis reports that SFN-YOLO focuses more sharply on poultry instances, produces more discrete and stable activations, and better handles illumination contrast, occlusion, and small targets. The authors interpret this as evidence that the global-local interactive loop suppresses invalid background responses while preserving object continuity.

6. Nomenclature, adjacent YOLO literature, limitations, and application scope

SFN-YOLO should be distinguished from similarly named YOLO variants that address different problem settings. “SF-YOLO” is a source-free domain adaptation framework for YOLO object detection rather than a poultry detector; it uses a teacher-student framework with a Target Augmentation Module and a Student Stabilisation Module, operates with unlabeled target data, and keeps YOLO inference unchanged (Varailhon et al., 2024). “YOLO LwF” is a self-distillation approach for continual object detection in YOLOv8 rather than a scale-aware fusion architecture; it addresses catastrophic forgetting in class incremental learning and combines distillation with replay memory and OCDM (Monte et al., 6 Mar 2025). By contrast, SFN-YOLO is a YOLOv8-based free-range poultry detector centered on scale-aware fusion and evaluated on the newly introduced M-SCOPE benchmark (Chen et al., 21 Sep 2025).

This distinction matters because the similar names can obscure the fact that the three systems target different axes of adaptation: source-free domain adaptation in SF-YOLO, continual object detection in YOLO LwF, and multiscale cluttered poultry detection in SFN-YOLO. A plausible implication is that architectural innovations, adaptation strategies, and evaluation metrics are not interchangeable across these works even when all are framed within the YOLO family.

The paper does not provide a formal limitations section, but several constraints are stated or implied. The main benchmark remains a single-class chicken detection setting. The dataset is new but not yet extremely large compared to generic detection corpora. The full model uses self-attention and may therefore be heavier than the tiniest YOLO variants, even though the reported efficiency remains high. The gains over YOLOv8s are modest in mAP, although they are paired with fewer parameters and stronger generalization. These points delimit the current evidence base without undermining the stated deployment focus.

Within that scope, the paper positions SFN-YOLO as a practical detection engine for flock counting, behavior monitoring, abnormality detection, farm automation, and robust deployment in free-range systems. Its reported strengths are real-time performance, lower parameter count, better small-object detection, more robust behavior under clutter and occlusion, and better cross-scene transfer (Chen et al., 21 Sep 2025).