SOD-YOLO: Enhanced UAV Small Object Detector

• The paper demonstrates a significant improvement in small object detection accuracy on VisDrone2019-DET, reporting a 36.1% increase in mAP50:95 over YOLOv8-m.
• It introduces three key modifications: an ASF mechanism for refined multi-scale feature fusion, a high-resolution P2 detection branch, and the use of Soft-NMS to preserve overlapping true positives.
• These enhancements enable more effective detection of tiny objects like pedestrians and distant vehicles in complex UAV imagery, albeit with increased computational demands.

SOD-YOLO is a YOLOv8-based object detector specialized for small object detection in UAV imagery, introduced in “SOD-YOLO: Enhancing YOLO-Based Detection of Small Objects in UAV Imagery” by Peijun Wang and Jinhua Zhao (Wang et al., 17 Jul 2025). It is designed for regimes in which targets such as pedestrians, bicycles, and distant vehicles occupy very few pixels, may be densely packed, and appear in cluttered aerial backgrounds. The method modifies the baseline detector at three points: it introduces an ASF mechanism in the neck to improve multi-scale feature fusion, adds a Small Object Detection Layer named P2 to exploit higher-resolution features, and replaces standard non-maximum suppression with Soft-NMS. On VisDrone2019-DET, the paper reports a 36.1% increase in $\mathrm{mAP}_{50:95}$ and a 20.6% increase in $\mathrm{mAP}_{50}$ relative to the baseline model (Wang et al., 17 Jul 2025).

1. Problem setting and motivation

SOD-YOLO addresses small object detection in UAV imagery, with VisDrone2019-DET as its primary benchmark (Wang et al., 17 Jul 2025). The paper frames this setting as difficult because UAV scenes contain many objects that occupy only a small fraction of the image, are often densely distributed, and are embedded in complex backgrounds. A central quantitative motivation is that, in VisDrone, over 75% of annotated objects occupy less than 0.1% of the image area (Wang et al., 17 Jul 2025).

Within that setting, the paper argues that standard YOLOv8 is constrained by three interacting factors. First, repeated downsampling weakens or removes the fine-grained spatial cues needed to discriminate tiny targets from background texture. Second, the default neck fusion strategy is not sufficiently selective for small-object-focused multi-scale aggregation. Third, standard post-processing can suppress true positives when multiple small detections overlap (Wang et al., 17 Jul 2025).

The proposed remedy is therefore not a wholesale redesign of the YOLO family, but a targeted reconfiguration of feature fusion, detection resolution, and post-processing. In that sense, SOD-YOLO is a task-specialized derivative of YOLOv8 rather than a new detector family.

2. Architectural design

The baseline described in the paper is YOLOv8-m, with a standard backbone, neck, and multi-scale detection head. The paper describes standard YOLOv8 as using Conv, C2f, and SPPF layers in the backbone, progressive feature fusion in the neck, and prediction heads at P3, P4, and P5 (Wang et al., 17 Jul 2025).

The first architectural change is the insertion of an ASF mechanism in the neck. In the paper, ASF is realized through two coupled components: Scale Sequence Feature Fusion, abbreviated ScalSeq, and an ASF attention refinement block. Two ScalSeq modules are inserted at critical stages of neck feature aggregation, and each receives three feature maps from P3, P4, and P5 (Wang et al., 17 Jul 2025).

The paper describes each ScalSeq module procedurally. A $1\times1$ convolution first unifies the channel dimensions of the three inputs. P4 and P5 are then upsampled with nearest-neighbor interpolation to the spatial size of P3. The aligned features are treated as a sequence along a new scale dimension, processed by a 3D convolution with kernel size $(1,1,1)$, followed by batch normalization and LeakyReLU, and then collapsed back to a standard 2D feature map by 3D max pooling with kernel size $(3,1,1)$ (Wang et al., 17 Jul 2025). This is more structured than ordinary channel concatenation because the fusion is explicitly organized across scale levels rather than only across channels.

After each ScalSeq block, the paper inserts an attention_model implementing ASF attention. This block takes two feature maps as input, applies channel attention to the first input feature map, adds the result to the second input feature map, and then applies local attention to the summed feature map (Wang et al., 17 Jul 2025). The stated purpose is to emphasize informative channels and spatial regions while suppressing background noise.

The second major modification is the Small Object Detection Layer named P2. The paper describes P2 as a higher-resolution detection branch added alongside the original P3–P5 heads (Wang et al., 17 Jul 2025). Its construction is summarized as follows: selected backbone feature maps are upsampled, concatenated with shallow-layer outputs, refined through a C2f module, passed through a ScalSeq module, and then used as an additional detection head. The final detector therefore predicts at four scales—P2, P3, P4, and P5—rather than the original three (Wang et al., 17 Jul 2025).

This higher-resolution branch is the paper’s explicit mechanism for preserving fine spatial detail. The design rationale is that very small objects may already be severely degraded at the coarser detection levels used by the baseline. The P2 head is introduced to retain edges, contours, and local structure that are otherwise weakened by repeated downsampling.

3. Post-processing and mathematical formulation

The third defining modification is the replacement of standard NMS with Soft-NMS (Wang et al., 17 Jul 2025). The paper presents this as an inference-time change intended to reduce the deletion of valid overlapping detections in dense UAV scenes.

The overlap measure used is standard IoU:

$\mathrm{IoU} = \frac{|A \cap B|}{|A \cup B|}.$

For standard NMS, the paper gives the score-update rule

$$S_i = \begin{cases} s_i & \mathrm{IoU}(A, B_i) < N_t \ 0 & \mathrm{IoU}(A, B_i) \geq N_t \end{cases}$$

where $S_i$ is the updated score of the $i$-th detection box, $s_i$ is its original score, $\mathrm{mAP}_{50}$0 is the highest-confidence detection, $\mathrm{mAP}_{50}$1 is another detection box, and $\mathrm{mAP}_{50}$2 is the IoU threshold (Wang et al., 17 Jul 2025).

For Soft-NMS, the paper replaces hard suppression with score decay:

$\mathrm{mAP}_{50}$3

This formulation is used to preserve overlapping true positives that standard NMS would eliminate outright (Wang et al., 17 Jul 2025).

The paper also provides intended formulas for precision and recall, while noting that the extracted LaTeX is malformed:

$\mathrm{mAP}_{50}$4

and

$\mathrm{mAP}_{50}$5

It further defines $\mathrm{mAP}_{50}$6 as mean average precision at IoU threshold $\mathrm{mAP}_{50}$7, and $\mathrm{mAP}_{50}$8 as mean average precision averaged across IoU thresholds from $\mathrm{mAP}_{50}$9 to $1\times1$0 (Wang et al., 17 Jul 2025).

By contrast, the paper does not provide explicit training-loss equations for classification, objectness, DFL, or box regression, even though the model is based on YOLOv8 (Wang et al., 17 Jul 2025). It likewise does not provide closed-form formulas for ASF attention or ScalSeq fusion beyond the architectural description. This omission is material for reproduction: the paper specifies the modules structurally, but not as a complete mathematical system.

4. Dataset, training protocol, and empirical results

The reported experiments are conducted on VisDrone2019-DET (Wang et al., 17 Jul 2025). The paper states that the dataset contains 10,209 static images and 261,908 video frames from 288 drone video clips. For the static detection set, it reports 6,471 training images, 548 validation images, and 1,610 testing images. The 10 object categories are pedestrian, people, bicycle, car, van, truck, tricycle, awning-tricycle, bus, and motor (Wang et al., 17 Jul 2025).

Training is reported on a single NVIDIA RTX 4090 using SGD with a cosine learning-rate schedule. The hyperparameters explicitly given are a linear warm-up to 0.005 in 3 epochs, weight decay 0.0005, momentum 0.937, 200 epochs, input image size $1\times1$1, and batch size 8 (Wang et al., 17 Jul 2025).

The main comparison table reports the following detector-level results on VisDrone2019-DET-val (Wang et al., 17 Jul 2025):

Model	$1\times1$2	$1\times1$3
YOLOv5-m	0.225	0.385
YOLOv7-m	0.265	0.472
YOLOv8-m	0.258	0.436
YOLOv9-gelan-c	0.305	0.489
YOLOv10-l	0.286	0.462
SOD-YOLO	0.351	0.526

The same table reports 22.6M parameters and 94.9G FLOPs for SOD-YOLO, compared with 25.8M parameters and 78.7G FLOPs for YOLOv8-m (Wang et al., 17 Jul 2025). Thus the paper’s headline gains over YOLOv8-m are absolute improvements of $1\times1$4 in $1\times1$5 and $1\times1$6 in $1\times1$7, corresponding to the reported relative increases of 36.1% and 20.6%.

The ablation study decomposes the contribution of the three modifications. Starting from the baseline at 0.258 $1\times1$8 and 0.436 $1\times1$9, adding ASF yields 0.265 and 0.440. Adding P2 on top of ASF yields 0.294 and 0.476. The full configuration with ASF, P2, and Soft-NMS is reported as 0.352 and 0.526 in the ablation table (Wang et al., 17 Jul 2025). This progression indicates that ASF contributes a modest gain, the P2 branch contributes a larger gain, and Soft-NMS contributes a further substantial increase.

The paper also presents qualitative comparisons, stating that SOD-YOLO detects more small pedestrians, better finds distant vehicles and motors, better handles partial occlusion, and improves discrimination among visually similar vehicle categories (Wang et al., 17 Jul 2025).

5. Relation to adjacent YOLO variants and nomenclature

The name “SOD-YOLO” belongs specifically to the UAV-oriented small-object detector described above (Wang et al., 17 Jul 2025). It should be distinguished from several similarly named but methodologically different YOLO derivatives.

A closely related name is “SOD-YOLOv8 -- Enhancing YOLOv8 for Small Object Detection in Traffic Scenes” (Khalili et al., 2024). That paper is also a small-object detector, but it targets traffic scenes rather than UAV imagery and modifies YOLOv8s through an efficient GFPN-inspired neck, a fourth detection layer, C2f-EMA modules, and PIoU loss. Its problem domain, baseline, and module composition are therefore distinct from SOD-YOLO (Khalili et al., 2024).

“YOLO-DS: Fine-Grained Feature Decoupling via Dual-Statistic Synergy Operator for Object Detection” is not a SOD paper in the naming sense: DS stands for Dual-Statistic Synergy rather than SOD (Huang et al., 26 Jan 2026). It is a general YOLOv8-based detector for heterogeneous-object response modeling on MS COCO, not a UAV small-object detector.

“SDD-YOLO: A Small-Target Detection Framework for Ground-to-Air Anti-UAV Surveillance with Edge-Efficient Deployment” is conceptually adjacent because it also targets tiny aerial targets and also adds a high-resolution P2 head (Chen et al., 26 Mar 2026). However, it is a YOLO26-derived anti-UAV detector with DFL-free and NMS-free inference, MuSGD, ProgLoss, and STAL, and is therefore a distinct design line.

Other similarly branded systems are further removed. “SatSplatYOLO” is a 3D Gaussian-splatting-based virtual-view ensemble wrapper around YOLOv5s for satellite component recognition, not a small-object UAV detector (Nguyen et al., 2024). “SSDA-YOLO” is a semi-supervised domain adaptive YOLOv5 framework for cross-domain object detection, not a small-object detector per se (Zhou et al., 2022). Earlier small-object-oriented or pyramid-enhanced YOLO variants such as “DC-SPP-YOLO” (Huang et al., 2019) and “YOLO-S” (Betti, 2022) are historically relevant as antecedents in feature reuse, pyramid pooling, and aerial small-target detection, but they do not define the SOD-YOLO method.

The resulting nomenclature is therefore narrower than the acronym might suggest. In the literature represented here, “SOD-YOLO” refers specifically to the YOLOv8-m-based UAV detector with ASF, P2, and Soft-NMS (Wang et al., 17 Jul 2025).

6. Limitations, reproducibility, and significance

The main limitation explicitly visible in the reported results is computational growth. FLOPs increase from 78.7G in YOLOv8-m to 94.9G in SOD-YOLO, and the paper acknowledges that this can reduce suitability for constrained real-time deployment despite the accuracy gains (Wang et al., 17 Jul 2025). Although the parameter count is reported as lower than the baseline, the added high-resolution branch and enhanced neck still increase the compute burden.

A second limitation is incomplete specification. The paper does not fully provide exact channel widths at each stage, a complete neck topology, explicit ASF equations, the augmentation recipe, Soft-NMS thresholds, or whether pretrained initialization is used (Wang et al., 17 Jul 2025). It also does not report FPS or latency, even though the method is motivated by practical UAV detection. Accordingly, exact reimplementation from the paper text alone is incomplete.

A third limitation is scope of validation. The paper’s empirical evidence is centered on VisDrone2019-DET, and it does not provide broader cross-dataset generalization studies, weather-robustness analysis, or edge-device deployment benchmarks (Wang et al., 17 Jul 2025). The qualitative examples support the intended use case, but the evidence remains domain-specific.

At the same time, the method is reproducibility-oriented in one important respect: the paper states that source code, hyper-parameters, and model weights are available at the linked GitHub repository (Wang et al., 17 Jul 2025). The ablation evidence also clarifies where the reported gains arise. A plausible implication is that, within this design, higher-resolution prediction through P2 and gentler suppression through Soft-NMS account for most of the improvement, while ASF contributes a smaller but positive refinement to multi-scale fusion.

In the broader evolution of YOLO-based small-object detection, SOD-YOLO occupies a recognizable design niche: it is a UAV-specific enhancement of YOLOv8 that intervenes simultaneously at feature fusion, detection resolution, and post-processing. Its significance lies less in proposing a new universal detection formalism than in demonstrating that these three intervention points, when combined in a domain-matched way, can materially improve detection of tiny aerial objects on VisDrone2019-DET (Wang et al., 17 Jul 2025).