MASF-YOLO: Scale-Adaptive UAV Object Detection

• The paper introduces MASF-YOLO, an augmented YOLOv11 framework incorporating a P2 detection head, MFAM, IEMA, and DASI to improve small object detection from UAV imagery.
• It employs multi-scale feature aggregation and attention mechanisms to efficiently combine fine spatial details with high-level semantics in cluttered scenes.
• Empirical evaluations on the VisDrone2019 benchmark demonstrate significant mAP improvements and reduced computational costs compared to larger YOLO variants.

MASF-YOLO (Multi-scale Context Aggregation and Scale-adaptive Fusion YOLO) is an object detection framework specifically designed for robust small object detection, particularly in imagery acquired from Unmanned Aerial Vehicle (UAV) perspectives. Developed as an augmentation of the YOLOv11 architecture, MASF-YOLO addresses the inherent difficulties in drone-based detection tasks—including the very small pixel proportion of targets, significant inter-object scale variations, and cluttered backgrounds—through a series of novel architectural modules for multi-scale context aggregation, attention, and scale-wise feature fusion (Lu et al., 25 Apr 2025).

1. Architectural Overview

MASF-YOLO is structured atop a YOLOv11 single-stage detection pipeline, introducing four major modifications:

1. Addition of a P2 Detection Head: A new detection head at the finest stride (stride 4, termed P2) is appended to the existing P3–P5 heads. This extension enables direct anchoring and prediction on the highest resolution feature maps, which is critical for small object recall.
2. Multi-scale Feature Aggregation Module (MFAM): Inserted preceding every C3 block in the backbone, MFAM fuses local and contextual cues from parallel multi-scale depthwise-separable convolutions, capturing a broader context at each scale with minimal additional computational cost.
3. Improved Efficient Multi-scale Attention Module (IEMA): Deployed after every downsampling in both backbone and neck, IEMA employs grouped, directional convolutions and cross-spatial reweighting to attenuate background and enhance salient object regions.
4. Dimension-Aware Selective Integration Module (DASI): Located at each fusion node in the neck, DASI dynamically fuses low-level (detailed) and high-level (semantic) features via channel-wise gating, balancing information from diverse scales.

The detection pipeline follows: Input → CBS stem → Backbone (C3+MFAM, CBS+IEMA) → Neck (FPN+PAN+skip+IEMA+DASI) → Detection Heads (P2–P5)

The diagram structure, as described in the original source, connects all these modules along the standard single-stage flow, preserving compatibility with YOLOv11 anchors, regression/classification branches, and loss functions (Lu et al., 25 Apr 2025).

2. Multi-scale Feature Aggregation Module (MFAM)

MFAM is designed to enhance the backbone's capability to capture multi-scale spatial context. For an input feature map $X\in\mathbb{R}^{C\times H\times W}$, MFAM computes four parallel branches:

• $Y_1 = \mathrm{DWConv}_{3\times3}(X)$
• $Y_2 = \mathrm{DWConv}_{5\times5}(X)$
• $$Y_3 = \mathrm{DWConv}_{7\times1}(\mathrm{DWConv}_{1\times7}(X))$$
• $$Y_4 = \mathrm{DWConv}_{9\times1}(\mathrm{DWConv}_{1\times9}(X))$$

These are summed with the input and projected:

$$Z = X + Y_1 + Y_2 + Y_3 + Y_4,\qquad W = \mathrm{Conv}_{1\times1}(Z)$$

This configuration enables MFAM to simultaneously aggregate fine and coarse spatial information, as larger kernels (decomposed to $\{1\times k, k\times 1\}$ paths for efficiency) yield broader receptive fields without substantial increases in FLOPs. This structure is interleaved throughout the backbone to aid in discriminating targets that may be only a few pixels in extent (Lu et al., 25 Apr 2025).

3. Improved Efficient Multi-scale Attention Module (IEMA)

IEMA is a channel-grouped, spatially-aware attention module intended to suppress background noise and emphasize target regions. Given input $X \in \mathbb{R}^{C \times H \times W}$, IEMA proceeds as follows:

• Grouping: $X$ is split into $G=16$ channel groups.
• Directional Convolutions: For each group $G_g$:
  • $U_g = \mathrm{DWConv}_{3\times3}(G_g)$
  • $V_g = \mathrm{DWConv}_{1\times5}(G_g)$
  • $W_g = \mathrm{DWConv}_{5\times1}(G_g)$
  • $I_g = G_g$
• The sum is passed through a sigmoid to obtain mask $M_g = \sigma(U_g + V_g + W_g + I_g)$, which reweights $G_g$. Concatenating all $\widetilde G_g = M_g\odot G_g$ reconstructs the channel dimension.

To further infuse cross-spatial information, averaged pooled feature vectors $Q$ and $K$ of shape $C \times 1$ are computed, and their scalar attention $A = \mathrm{softmax}(Q^\top K)$ modulates the feature map. Final output is:

$$O = \sigma(\mathrm{Conv}_{1\times 1}(A\widetilde G))\odot\widetilde G + \widetilde G$$

This layered, directionally-aware attention mechanism augments detector precision, particularly for targets that are spatially small and surrounded by complex backgrounds (Lu et al., 25 Apr 2025).

4. Dimension-Aware Selective Integration Module (DASI)

DASI adaptively combines low-, mid-, and high-level features in the neck using channel-wise attention and partitioned gating. Three feature maps $F_\ell$ (low-level), $F_c$ (current), and $F_h$ (high-level) are aligned to the same spatial and channel dimensions and partitioned along the channel axis:

$$F_* = [F_*^1, \ldots, F_*^4], \quad F_*^i \in \mathbb{R}^{\frac{C}{4} \times H \times W}$$

For each partition $i$, a gating weight $\alpha_i = \sigma(F_c^i)$ is learned. Fusion at each partition is:

$$R_i = \alpha_i \odot F_\ell^i + (1-\alpha_i) \odot F_h^i$$

The fused output is:

$$F_{\mathrm{out}} = \mathrm{Conv}_{1\times1}([R_1,\ldots,R_4]) + F_c$$

DASI's partitioned blending enables channelwise, instance-specific filtering between spatial detail and semantics, critical for scale-adaptive detection in UAV scenes where both foreground scale and clutter background co-occur (Lu et al., 25 Apr 2025).

5. Small Object Detection Enhancements and Detection Heads

The addition of the P2 (stride-4) detection head in MASF-YOLO distinguishes it from conventional YOLO-style architectures. By directly regressing and classifying bounding boxes at the input's highest spatial resolution, P2 significantly elevates the detection rate for diminutive targets common in aerial imagery. This is further supported by the skip connections fused into the neck, which preserve early spatial details throughout the detection pathway.

No changes are made to anchor definitions, loss functions, or heads' prediction format; the modules are fully compatible with standard YOLOv11 detection protocols, using the same objectness, bounding-box regression, and classification branches.

6. Empirical Evaluation and Ablation

Experimental evaluation on the VisDrone2019 benchmark demonstrates the efficacy of the proposed modules:

Model	[email protected] (val)	[email protected]:0.95 (val)	Params	GFLOPs
YOLOv11-s	44.6 %	29.4 %	9.42M	21.3
MASF-YOLO-s	49.2 %	32.9 %	12.05M	44.3
YOLOv11-m	47.8 %	32.2 %	20.04M	67.7

MASF-YOLO-s achieves sizable improvements over YOLOv11-s (+4.6% [email protected] and +3.5% [email protected]:0.95) and outperforms even larger models (YOLOv11-m) despite using only about 60% of the parameters and 65% of the computational budget. Comparative tests on the VisDrone2019 validation set show that MASF-YOLO-s attains the highest reported mAP among a set of single- and two-stage detectors, including EfficientDet-D0, TPH-YOLOv5-s, and YOLOv8-m.

Ablation analysis indicates that each constituent module (P2, MFAM, skip-fusion, IEMA, DASI) incrementally improves detection performance, with the greatest absolute boost from the initial integration of MFAM and the P2 head (Lu et al., 25 Apr 2025).

7. Significance and Applications

MASF-YOLO demonstrates that multi-scale context aggregation, scale-adaptive channel-wise fusion, and fine-resolution pyramidal regression substantially advance the practical viability of real-time, small object detection under the unique constraints of UAV viewpoints. This architecture is particularly pertinent to scenarios such as aerial surveillance, traffic monitoring, and remote sensing, where accurate localization and classification of dense, small-scale targets in cluttered environments are required.

A plausible implication is that further gains may be possible by integrating temporal consistency or leveraging transformer-style long-range dependencies in similar modular architectures, though such extensions are not addressed in the referenced design. MASF-YOLO exemplifies efficient, modular improvements to the standard YOLO paradigm and establishes a new performance baseline for single-stage, drone-view object detection (Lu et al., 25 Apr 2025).