---
title: 'MS-YOLO: A Multiscale YOLO Approach'
url: https://www.emergentmind.com/topics/ms-yolo
type: topic
---

# MS-YOLO: A Multiscale YOLO Approach

Searching arXiv for papers on “MS-YOLO” and closely related variants to ground the article in current literature.
MS-YOLO denotes a small family of YOLO-derived detectors rather than a single canonical architecture. In current arXiv usage, the name has been applied to at least three distinct systems: the MultiScale Domain Adaptive YOLO built on YOLOv4 for cross-domain object detection, a YOLOv8-based infrared detector for edge deployment that combines MobileNetV4 Small with SlideLoss, and a YOLOv11-based blood cell detector centered on multiscale feature extraction and fusion. Closely related lines, notably YOLO-MS and YOLO-MST, are explicitly framed as multiscale YOLO approaches and help define the broader technical meaning of “MS-YOLO” as a design orientation toward multiscale representation learning, difficult operating conditions, and real-time detection [2106.01483] [2509.21696] [2506.03972] [2308.05480] [2412.19878].

## 1. Terminological scope and disambiguation

The term “MS-YOLO” is best understood as polysemous. In some papers, “MS” denotes **multiscale domain adaptation**; in others it denotes a **multi-scale model** tailored to a specific sensing modality or application. A recurring misconception is that MS-YOLO identifies one standardized detector family analogous to YOLOv5 or YOLOv8. The literature instead uses the label for multiple, architecturally different systems that share a commitment to multiscale feature processing [2106.01483] [2202.03527] [2509.21696] [2506.03972].

| Work | Base model | Reported focus |
|---|---|---|
| "Multiscale Domain Adaptive YOLO" [2106.01483] | YOLOv4 | Cross-domain object detection |
| "Integrated Multiscale Domain Adaptive YOLO" [2202.03527] | YOLOv4 | PFR, UC, and Integrated DAN architectures |
| "MS-YOLO: Infrared Object Detection for Edge Deployment via MobileNetV4 and SlideLoss" [2509.21696] | YOLOv8 | Thermal urban edge deployment |
| "MS-YOLO: A Multi-Scale Model for Accurate and Efficient Blood Cell Detection" [2506.03972] | YOLOv11 | Complete blood cell detection |
| "YOLO-MS: Rethinking Multi-Scale Representation Learning for Real-time Object Detection" [2308.05480] | YOLO-style detector | MS-Block and HKS |
| "YOLO-MST" [2412.19878] | YOLOv5 | Infrared small target detection |

Across these works, “multiscale” refers to different technical mechanisms. In MS-DAYOLO it means attaching adversarial domain classifiers to multiple backbone scales; in the thermal and microscopy variants it refers to multiscale feature extraction, fusion, and loss reweighting; in YOLO-MS it is encoded directly inside the basic block and stage-wise kernel selection; and in YOLO-MST it combines super-resolution, multiscale aggregation, and a dynamic multiscale head [2202.03527] [2509.21696] [2506.03972] [2308.05480] [2412.19878].

## 2. MS-DAYOLO: multiscale domain adaptation in YOLOv4

The earliest MS-YOLO line is MS-DAYOLO, introduced as a domain-adaptive extension of YOLOv4 for cross-domain object detection, especially for autonomous driving under weather and sensor shifts. The core design taps three backbone feature maps, \(F_1\), \(F_2\), and \(F_3\), before the neck and attaches three domain adaptation paths, each containing a Gradient Reversal Layer and a domain classifier. Detection is optimized on labeled source images, while the Domain Adaptive Network is trained to distinguish source from target; the backbone, through gradient reversal, is trained to confuse the domain classifiers and thereby learn domain-invariant features [2106.01483].

For a training image \(i\), domain label \(t_i \in \{0,1\}\), and per-location domain probability \(p_i^{(x,y)}\), the domain-classification loss is
$$
\mathcal{L}_{dc} = -\sum_{i,x,y} \left[ t_i \ln p_i^{(x,y)} + (1 - t_i) \ln (1 - p_i^{(x,y)}) \right].
$$
The backbone is trained with
$$
\mathcal{L}_t = \mathcal{L}_{det} + \lambda \mathcal{L}_{dc},
$$
with \(\lambda = 0.1\), while the sign reversal is handled by the GRL. Training batches contain 64 images, split into 32 labeled source images and 32 unlabeled target images. At inference, the DAN and GRLs are removed, so deployment uses standard YOLOv4 inference with domain-adapted weights [2106.01483].

The 2021 paper reports that, on a clear-to-foggy Cityscapes protocol using five classes, baseline YOLOv4 achieves a mAP of 32.77, full multiscale MS-DAYOLO reaches 41.05, and an oracle YOLOv4 trained directly on foggy data reaches 45.24. Single-scale and two-scale adaptations improve over baseline, but the best result is obtained when all three scales are adapted simultaneously, supporting the claim that low-, mid-, and high-level domain shifts should be handled jointly rather than only at the deepest feature layer [2106.01483].

The 2022 extension deepens this line by introducing three alternative Domain Adaptation Network architectures: **Progressive Feature Reduction (PFR)**, **Unified Classifier (UC)**, and an **Integrated** architecture that combines both ideas. PFR replaces abrupt two-layer reduction with four convolutional stages for \(F_1\) and \(F_2\), and five for \(F_3\), progressively reducing channels to a single-channel domain logit map. UC addresses the inconsistency of scale-specific domain classifiers by aligning and concatenating multiscale features into one representation and using a shared domain classifier. The Integrated architecture combines multiscale aggregation with progressive reduction and is the strongest of the three [2202.03527].

Empirically, the 2022 paper reports that on KITTI \(\rightarrow\) Cityscapes, plain YOLOv4 achieves 44.5 AP for the car class, whereas Integrated MS-DAYOLO reaches 47.6 AP while running at about 48.2 FPS on a GTX 1080 Ti; Faster R-CNN-based domain-adaptive baselines are reported around 3.6–6.2 FPS. On the clear \(\rightarrow\) foggy Cityscapes experiment, plain YOLOv4 reaches mAP 31.0 and Integrated MS-DAYOLO reaches 41.5, while on sunny \(\rightarrow\) rainy Waymo, YOLOv4 reaches mAP 47.0 and Integrated reaches 48.5. These results establish MS-DAYOLO as a real-time adversarial domain-adaptive detector rather than a generic multiscale backbone [2202.03527].

## 3. MS-YOLO for infrared urban detection and edge deployment

A different meaning of MS-YOLO appears in "MS-YOLO: Infrared Object Detection for Edge Deployment via MobileNetV4 and SlideLoss" [2509.21696]. Here the detector is a modified YOLOv8-based system for thermal urban scenes and resource-constrained deployment. The baseline selection process evaluates YOLOv5n, YOLOv8n, YOLOv9t, YOLOv10n, and YOLOv11n on FLIR ADAS V2, and the retrained YOLOv8n is chosen because it achieves the best recall, the best mAP@0.5, and the best mAP@0.5:0.95 among the listed nano/tiny baselines, with 6.8 GFLOPs [2509.21696].

MS-YOLO then replaces YOLOv8’s CSPDarknet backbone with **MobileNetV4 Small**, built from Conv2D blocks and **Universal Inverted Bottleneck** stacks, while preserving the YOLOv8 PAN/FPN-style neck and three-scale detection heads. The input is \(640 \times 640 \times 1\) thermal imagery. The second modification is **SlideLoss**, an IoU-weighted loss reweighting strategy adapted from occlusion-aware face detection. With \(x\) the IoU between prediction and ground truth, and \(\mu\) the average IoU over training boxes, the weighting function is
$$
f(x) = \begin{cases}
1, & x \le \mu - 0.1, \\
e^{1-\mu}, & \mu - 0.1 < x < \mu, \\
e^{1-x}, & x \ge \mu .
\end{cases}
$$
The total weighted loss is written generically as
$$
L_{\text{MS-YOLO}} = \sum_i f(x_i)\left(L_{\text{cls},i}+L_{\text{obj},i}+L_{\text{box},i}\right).
$$
This formulation is intended to emphasize hard, under-represented, and occluded samples without changing inference cost [2509.21696].

The reported operating domain is the thermal-only subset of FLIR ADAS V2, filtered to nine classes: person, bike, car, motor, bus, truck, light, hydrant, and sign. The selected subset contains 15,094 images and 244,618 instances, split into 10,474 training images, 3,493 test images, and 1,127 validation images. Training uses a single NVIDIA Tesla V100 SXM2, 200 epochs, \(640\times640\) inputs, and batch size 128 [2509.21696].

Quantitatively, baseline YOLOv8n* reports Precision 0.624, Recall 0.463, mAP@0.5 0.517, mAP@0.5:0.95 0.318, and 6.8 GFLOPs. MobileNetV4 alone reduces GFLOPs to 6.7 and raises Precision to 0.643, but lowers Recall and mAP. SlideLoss alone yields Precision 0.636, Recall 0.459, mAP@0.5 0.505, mAP@0.5:0.95 0.300, and 6.8 GFLOPs. The final MS-YOLO reports Precision 0.649, Recall 0.430, mAP@0.5 0.484, mAP@0.5:0.95 0.282, and 6.7 GFLOPs. The paper therefore positions the model as a precision-oriented, edge-friendly thermal detector rather than the highest-mAP configuration on FLIR ADAS V2 [2509.21696].

The class-wise results underscore the imbalance problem the model is meant to address. In the validation subset, person has 4,309 instances and car 7,128, while motorcycle has 55, truck 46, and fire hydrant 94. Reported MS-YOLO class-wise performance includes car with Precision 0.810, Recall 0.729, and mAP@0.5 0.814; truck with Precision 0.191, Recall 0.261, and mAP@0.5 0.0965; fire hydrant with Precision 0.676, Recall 0.177, and mAP@0.5 0.264; and motorcycle with Precision 0.761, Recall 0.579, and mAP@0.5 0.607. This suggests that SlideLoss can help hard or rare categories, but does not eliminate severe scarcity and inter-class confusion [2509.21696].

## 4. MS-YOLO for complete blood cell detection

The 2025 paper "MS-YOLO: A Multi-Scale Model for Accurate and Efficient Blood Cell Detection" [2506.03972] uses the same label for a specialized YOLOv11-based microscopy detector. Its target problem is complete blood cell detection under large scale disparity, dense overlap, and small-object sensitivity, especially for platelets. The model preserves the global YOLOv11 pipeline but replaces key modules in the backbone and neck with three new components: **MS-DRM**, **DCFEM**, and **LADS** [2506.03972].

**MS-DRM** replaces the bottleneck layers in YOLOv11’s C3K2 modules. It is derived from the Dilation-Wise Residual module and combines region residualization with channel-grouped depthwise separable convolutions at different dilation rates. With input \(X\), region residualization produces \(X_r\); channel groups are then processed by depthwise dilated convolutions and fused:
$$
Y^{(g)} = \mathrm{DWConv}_{3\times3}\left(X_r^{(g)}; \text{dilation}=r_g\right), \qquad
Z = X + \mathrm{BN}(\mathrm{Conv}_{1\times1}(\mathrm{Concat}_g\, Y^{(g)})).
$$
The motivation is to enlarge and diversify receptive fields without a prohibitive parameter increase, thereby improving discriminability across RBCs, WBCs, and tiny platelets [2506.03972].

**DCFEM** is inserted at the front-end of the detection head and performs bidirectional cross-path fusion between backbone features and neck features. After channel alignment and concatenation, it computes local and global branches, generates dynamic local/global weights, and fuses them according to
$$
F_{\text{fused}} = W_{\text{local}} \odot F_{\text{local}} + W_{\text{global}} \odot F_{\text{global}}.
$$
This is followed by attention calibration and residual feedback to the backbone and neck streams. The paper reports that DCFEM alone contributes \(+0.8\%\) mAP@0.5 and \(+1.5\%\) mAP@0.5:0.95 over the YOLOv11-n baseline, the strongest single-module gain among the three additions [2506.03972].

**LADS** replaces some standard downsampling modules in the backbone and neck. It uses a parallel design: one branch learns adaptive spatial weights over each \(2\times2\) region, and the other performs group convolution with stride 2. The downsampled output is a weighted aggregation,
$$
Y[u,v] = \sum_{i,j} w_{ij}(u,v)\, X_{ij}(u,v),
$$
intended to preserve fine-grained signals relevant to tiny platelets while keeping complexity moderate. In ablation, LADS adds \(+0.2\%\) mAP@0.5 and \(+0.5\%\) mAP@0.5:0.95 over baseline [2506.03972].

The reported training setup uses the CBC dataset with 360 microscopy images and three classes—RBC, WBC, and platelet—split into 240 train, 60 validation, and 60 test images. Images are resized to \(640\times640\), augmented with mosaic, random horizontal flip, and HSV perturbation, and trained from scratch with AdamW, learning rate 0.001, cosine annealing over 300 epochs, batch size 32, and no pretrained weights. A supplementary evaluation is performed on WBCDD, containing 684 images and five leukocyte subtypes [2506.03972].

Performance is reported against YOLOv11-n and a range of YOLO, DETR, and blood-cell-specific baselines. On CBC, YOLOv11-n reports mAP@0.5 95.8, mAP@0.5:0.95 67.9, AP\(_{\text{RBC}}\) 94.8, AP\(_{\text{WBC}}\) 99.5, AP\(_{\text{Platelet}}\) 93.2, Precision 83.3, and Recall 90.8. The full MS-YOLO reports mAP@0.5 97.4, mAP@0.5:0.95 69.9, AP\(_{\text{RBC}}\) 95.4, AP\(_{\text{WBC}}\) 99.5, AP\(_{\text{Platelet}}\) 97.3, Precision 86.4, and Recall 92.3. The platelet gain of \(+4.1\%\) AP is the most prominent class-specific improvement. The final model has 3.1M parameters, 8.7 GFLOPs, 6.3 MB weights, and 66.1 FPS on CBC. On WBCDD, it reports mAP@0.5 95.3 and the highest mAP@0.5:0.95 at 81.2, supporting the paper’s claim of robust generalization [2506.03972].

## 5. Related multiscale YOLO families

Two adjacent lines help clarify how the broader community has used the multiscale-YOLO idea even when the exact title differs.

"YOLO-MS: Rethinking Multi-Scale Representation Learning for Real-time Object Detection" [2308.05480] is a family of real-time one-stage detectors that often functions as a reference point for “MS-YOLO” in the generic sense. Its core innovation is the **MS-Block**, a hierarchical multi-branch block inspired by Res2Net, using inverted bottlenecks and depth-wise large kernels. Given split features \(X_i\), the branch outputs follow
$$
Y_i =
\begin{cases}
X_i, & i=1, \\
IB_{k\times k}(Y_{i-1}+X_i), & i>1.
\end{cases}
$$
This local multiscale design is paired with **Heterogeneous Kernel Selection**, which sets stage-wise depth-wise kernel sizes to \([3,5,7,9]\) in the backbone and \(3\times3\) in the neck. On MS COCO, YOLO-MS-XS reports 43.4 AP with 4.54M parameters, 8.74G FLOPs, and 7.6 ms latency; YOLO-MS-S reports 46.2 AP; and YOLO-MS reports 51.0 AP. As a plug-and-play module, it raises YOLOv8-n from 37.2 AP to 40.3 AP with fewer parameters and the same 4.4G FLOPs, and raises YOLOv6-tiny from 41.0 AP to 43.5 AP while reducing both parameters and FLOPs [2308.05480].

"YOLO-MST: Multiscale deep learning method for infrared small target detection based on super-resolution and YOLO" [2412.19878] applies the multiscale YOLO idea to infrared small target detection. Its pipeline combines **Real-ESRGAN** \(\times4\) super-resolution, a backbone **MSFA** module that replaces SPPF with multi-dilated feature aggregation, a neck optimized for small and medium scales, and a **DyHead** dynamic detection head with scale-aware, spatial, and task-channel attention. On SIRST it reports mAP@0.5 96.4, mAP@0.5:0.95 47.5, Precision 95.8, and Recall 91.8; on IRIS it reports mAP@0.5 99.5 and FPS 52.7. Ablation attributes the largest single gain to super-resolution, which raises SIRST mAP@0.5 from 69.3 to 86.1 over baseline YOLOv5s, while the full stack reaches 96.4 [2412.19878].

These two works show that outside domain adaptation and task-specific nomenclature, multiscale YOLO research has moved in two complementary directions: richer multiscale representation learning inside the backbone and neck, and explicit small-target optimization via super-resolution, dilated aggregation, and dynamic detection heads. A plausible implication is that “MS-YOLO” has become a shorthand for a design philosophy rather than a uniquely identifiable architecture [2308.05480] [2412.19878].

## 6. Comparative interpretation, recurring motifs, and open issues

Across the literature, several motifs recur. First, multiscale processing is consistently pushed closer to the feature extractor rather than left entirely to the FPN or PAN stage. MS-DAYOLO attaches adaptation paths to \(F_1\), \(F_2\), and \(F_3\); the blood-cell model replaces C3K2 with MS-DRM and inserts DCFEM before the head; YOLO-MS redesigns the basic block itself; and YOLO-MST replaces SPPF, alters the neck, and adds DyHead [2106.01483] [2506.03972] [2308.05480] [2412.19878].

Second, several variants separate **training-time complexity** from **deployment-time cost**. In MS-DAYOLO, DAN and GRLs are removed for inference, preserving YOLOv4 runtime. In the thermal edge model, SlideLoss changes only the training objective, so MobileNetV4 Small accounts for the deployment-side efficiency while the loss reweighting is computationally free at inference. This suggests a recurring engineering strategy: reserve additional complexity for supervision and adaptation when runtime budgets are stringent [2202.03527] [2509.21696].

Third, the published systems are highly application-specific. The domain-adaptive line is optimized for autonomous driving weather and camera shifts; the MobileNetV4-SlideLoss model is calibrated to thermal urban scenes with FLIR ADAS V2; the YOLOv11 variant is tailored to microscopy with RBC, WBC, and platelet scale disparities. This suggests that MS-YOLO, as used in the literature, is better viewed as a pattern of multiscale specialization than as a universally transferable detector family [2509.21696] [2506.03972].

The limitations reported by the papers are correspondingly heterogeneous. MS-DAYOLO introduces training overhead and depends on the choice of \(\lambda\), with evaluation focused mainly on weather-related domain shifts [2106.01483] [2202.03527]. The thermal MS-YOLO improves precision but yields lower recall and lower mAP than retrained YOLOv8n* on FLIR ADAS V2, and its domain generalization beyond a single thermal dataset is not tested [2509.21696]. The blood-cell MS-YOLO remains sensitive to limited data and rare morphologies, and its deployment on constrained edge hardware is not explicitly benchmarked despite its lightweight design [2506.03972]. YOLO-MS, while accurate, is slightly slower than some ultra-fast detectors because large-kernel convolutions and hierarchical multi-branch structures still have latency cost [2308.05480]. YOLO-MST depends heavily on super-resolution and is specialized to infrared small targets rather than general object detection [2412.19878].

Taken together, these works indicate that the most stable meaning of MS-YOLO is not a single paper-defined model but a research direction centered on multiscale feature handling under adverse conditions: domain shift, small targets, occlusion, class imbalance, or overlapping structures. In that sense, the encyclopedia-level definition of MS-YOLO is necessarily plural.

Source: https://www.emergentmind.com/topics/ms-yolo