Global Multi-scale Guidance Module (GMGM)

• GMGM is a decoder-side module that extracts global semantic context from shallow encoder features to generate guidance maps for structured building topology refinement.
• It employs global average pooling and lightweight cross-scale attention to derive adaptive weights and reproject features across multiple resolutions.
• Empirical evidence shows that incorporating GMGM enhances IoU and F1 scores under low-light and hazy conditions by anchoring global semantic priors.

Global Multi-scale Guidance Module (GMGM) denotes a module class in which globally aggregated semantic context is converted into scale-aligned guidance for feature refinement or decoding. In the literature provided here, the term is most explicitly instantiated in HaLoBuild-Net, where GMGM aggregates three-scale shallow encoder features, derives global descriptors through global average pooling, predicts adaptive cross-scale weights with two $1\times1$ convolutions and a ReLU, fuses the weighted features into a high-resolution global representation, and emits three guidance features that are concatenated into corresponding decoder stages to provide global semantic constraints and anchor building topologies under hazy and low-light conditions (Sang et al., 16 Apr 2026). Beyond that explicit usage, closely related mechanisms appear in multiple papers under different names, typically separating global context generation from multi-scale guided fusion rather than packaging both functions under one module label (Tao et al., 4 Mar 2026).

1. Definition and functional scope

In its explicit HaLoBuild-Net formulation, GMGM is a decoder-side guidance-generation branch built from encoder features. It receives the three shallow encoder features $$\{F_{\mathrm{enc1}}, F_{\mathrm{enc2}}, F_{\mathrm{enc3}}\}$$, produces three guidance features $$\{G_{\mathrm{enc1}}, G_{\mathrm{enc2}}, G_{\mathrm{enc3}}\}$$, and integrates those guidance features into decoder stages via concatenation (Sang et al., 16 Apr 2026). Its stated purpose is to stabilize building extraction when haze or low light makes local evidence unreliable, supplying a scene-level structural prior that improves completeness, positional consistency, and structural continuity.

Within that paper, GMGM is introduced to address failure modes that arise when local receptive fields become unreliable under meteorological degradation. The module is described as providing “global semantic constraints” and “explicit spatial priors to anchor building topologies,” with the practical effects framed in terms of preserving overall building completeness, preventing internal voids in large structures, and preventing small buildings from being overwhelmed by noise (Sang et al., 16 Apr 2026). This makes GMGM neither a restoration head nor a final-logit postprocessor. It is an internal feature-level guidance mechanism.

A broader reading across the cited literature suggests that GMGM is best understood as a functional pattern rather than a universally standardized primitive. A plausible abstraction is a module that first extracts global semantic or contextual priors and then redistributes them across multiple scales through concatenation, attention, gating, semantic flows, or pixel-wise weighting. That interpretation fits the explicit HaLoBuild-Net design and also fits several adjacent architectures that implement highly analogous behavior without using the same name.

2. Canonical formulation in HaLoBuild-Net

The most concrete GMGM definition in the supplied papers is the one given in HaLoBuild-Net (Sang et al., 16 Apr 2026). The module begins by applying global average pooling to each shallow encoder feature and concatenating the resulting descriptors:

$$V_{\mathrm{agg}} = \mathrm{Concat}_c \left( \{ \mathrm{GAP}(F_{\mathrm{enc}i}) \}_{i=1}^{3} \right)$$

This pooled descriptor is then passed through a lightweight cross-scale attention network consisting of one $1\times1$ convolution, a ReLU, and another $1\times1$ convolution, after which the result is split into three scale weights:

$$\{w_1, w_2, w_3\} = \mathrm{Split}(\mathrm{Conv}_{1 \times 1}(\delta(\mathrm{Conv}_{1 \times 1}(V_{\mathrm{agg}}))))$$

The three shallow features are reweighted and spatially aligned to a common high-resolution grid. The highest-resolution feature remains at its native scale, while the other two are upsampled by bilinear interpolation by factors $2\times$ and $4\times$. The weighted aligned features are then concatenated:

$$F_{\mathrm{global}} = \mathrm{Concat}_c \big( w_1 F_{\mathrm{enc1}}, U(w_2 F_{\mathrm{enc2}}), U(w_3 F_{\mathrm{enc3}}) \big)$$

Finally, $$\{F_{\mathrm{enc1}}, F_{\mathrm{enc2}}, F_{\mathrm{enc3}}\}$$0 is projected into three decoder-facing guidance tensors through three parallel ConvBNReLU branches:

$$\{F_{\mathrm{enc1}}, F_{\mathrm{enc2}}, F_{\mathrm{enc3}}\}$$1

$$\{F_{\mathrm{enc1}}, F_{\mathrm{enc2}}, F_{\mathrm{enc3}}\}$$2

Here $$\{F_{\mathrm{enc1}}, F_{\mathrm{enc2}}, F_{\mathrm{enc3}}\}$$3 denotes downsampling back to the original scale of the $$\{F_{\mathrm{enc1}}, F_{\mathrm{enc2}}, F_{\mathrm{enc3}}\}$$4-th encoder feature. The module therefore constructs one shared globally aggregated semantic representation and then reprojects it into stage-specific guidance maps.

This formulation clarifies the meanings of the three terms in the module name. “Global” refers to image-level context obtained by global average pooling over each shallow feature map. “Multi-scale” refers to the joint use of three encoder scales and their adaptive weighting. “Guidance” refers to the generation of decoder-side semantic priors that are concatenated into corresponding decoding stages (Sang et al., 16 Apr 2026). The paper explicitly distinguishes this role from the roles of its companion modules: SFFM performs spatial-frequency feature restoration, while MGFM performs bidirectional semantic-spatial calibration.

3. Architectural placement and interaction with the decoder

GMGM in HaLoBuild-Net operates on encoder features but is architecturally positioned as a decoder guidance branch. The LWGANet-L2 encoder produces four feature maps at resolutions $$\{F_{\mathrm{enc1}}, F_{\mathrm{enc2}}, F_{\mathrm{enc3}}\}$$5, $$\{F_{\mathrm{enc1}}, F_{\mathrm{enc2}}, F_{\mathrm{enc3}}\}$$6, $$\{F_{\mathrm{enc1}}, F_{\mathrm{enc2}}, F_{\mathrm{enc3}}\}$$7, and $$\{F_{\mathrm{enc1}}, F_{\mathrm{enc2}}, F_{\mathrm{enc3}}\}$$8 of the input image, and GMGM consumes the first three of these, not the deepest $$\{F_{\mathrm{enc1}}, F_{\mathrm{enc2}}, F_{\mathrm{enc3}}\}$$9 feature (Sang et al., 16 Apr 2026). This is important because the module is explicitly described as aggregating “multi-scale shallow features,” which preserves localization cues while still allowing global semantic planning.

The output guidance features $$\{G_{\mathrm{enc1}}, G_{\mathrm{enc2}}, G_{\mathrm{enc3}}\}$$0 are integrated into the features of each decoder stage via concatenation. The direction of information flow is therefore from globally aggregated shallow encoder semantics to decoder-stage refinement. GMGM does not act on segmentation logits and does not provide a separate supervisory objective. It is trained indirectly through the network’s segmentation loss, which combines Soft Cross-Entropy and Dice loss (Sang et al., 16 Apr 2026).

Within HaLoBuild-Net, the division of labor among decoder-side modules is explicit. SFFM mitigates meteorological interference on building features by coupling large receptive field attention with frequency-aware channel reweighting guided by stable low-frequency anchors. MGFM performs bidirectional semantic-spatial calibration to suppress shallow noise and sharpen weather-induced blurred boundaries. GMGM contributes a distinct function: global topological anchoring. The paper’s own characterization is that SFFM improves feature quality, MGFM improves cross-level fusion, and GMGM improves the global structural correctness of the decoded building map (Sang et al., 16 Apr 2026).

This division is significant because it separates semantic planning from restoration and boundary sharpening. GMGM is therefore not merely another fusion block. It is the module that supplies an explicit global prior for where buildings should be and how their overall structure should be organized when local evidence is degraded.

4. Related formulations in adjacent literature

Outside HaLoBuild-Net, several papers implement closely related mechanisms under different names. In these works, the GMGM concept is distributed across multiple submodules rather than instantiated as a single named block.

Paper	Closest GMGM-equivalent	Core mechanism
"Small Object Detection in Complex Backgrounds with Multi-Scale Attention and Global Relation Modeling" (Tao et al., 4 Mar 2026)	GRM + CSHA	Global semantic priors on $$\{G_{\mathrm{enc1}}, G_{\mathrm{enc2}}, G_{\mathrm{enc3}}\}$$1 followed by sparse aligned fusion across $$\{G_{\mathrm{enc1}}, G_{\mathrm{enc2}}, G_{\mathrm{enc3}}\}$$2
"FSGNet: A Frequency-Aware and Semantic Guidance Network for Infrared Small Target Detection" (Zhang et al., 26 Mar 2026)	GPM + GSGFs/GSGM	Bottleneck global semantic tensor propagated to multiple decoder stages
"MGDFIS: Multi-scale Global-detail Feature Integration Strategy for Small Object Detection" (Wang et al., 15 Jun 2025)	GDIM, especially GMM + DMM	Directional global mixing followed by detail-preserving local enhancement
"Cross-Layer Feature Self-Attention Module for Multi-Scale Object Detection" (Xie et al., 16 Oct 2025)	CFSAM	Joint cross-layer self-attention over flattened multi-scale features
"Speech Emotion Recognition with Global-Aware Fusion on Multi-scale Feature Representation" (Zhu et al., 2022)	Multi-scale block + global-aware fusion module	Multi-scale convolutional representation followed by gMLP-style global fusion

In the small-object detector of (Tao et al., 4 Mar 2026), the paper explicitly states that it does not define a module called GMGM, but that the combination of the Global Relation Modeling module and the Cross-Scale Hybrid Attention module is a close functional equivalent. GRM generates global semantic priors by applying multi-head self-attention to flattened $$\{G_{\mathrm{enc1}}, G_{\mathrm{enc2}}, G_{\mathrm{enc3}}\}$$3, while CSHA uses those semantically enriched features to guide sparse, aligned fusion across $$\{G_{\mathrm{enc1}}, G_{\mathrm{enc2}}, G_{\mathrm{enc3}}\}$$4, $$\{G_{\mathrm{enc1}}, G_{\mathrm{enc2}}, G_{\mathrm{enc3}}\}$$5, and $$\{G_{\mathrm{enc1}}, G_{\mathrm{enc2}}, G_{\mathrm{enc3}}\}$$6. The paper explicitly describes this as separating “global guidance generation” from “multi-scale guided interaction.”

FSGNet presents a narrower but still closely analogous case. Its deepest-layer Global Pooling Module builds a four-branch hierarchical pooling structure over the bottleneck feature, fuses the outputs into $$\{G_{\mathrm{enc1}}, G_{\mathrm{enc2}}, G_{\mathrm{enc3}}\}$$7, and then propagates that semantic tensor upward through multiple Global Semantic Guidance Flows into decoder stages (Zhang et al., 26 Mar 2026). This is GMGM-like in the sense of generating a globally informed semantic source and distributing it across scales, although the paper’s own emphasis is on semantic preservation in U-Net decoding rather than on a general global multi-scale guidance abstraction.

MGDFIS is closer to a neck-level GMGM interpretation. Its Global-detail Integration Module first aggregates two aligned features, then applies the Global Mixing Module to mine long-range dependencies via directional slicing and reassembly, and finally applies the Detail Mixing Module to preserve orientation-sensitive local detail through $$\{G_{\mathrm{enc1}}, G_{\mathrm{enc2}}, G_{\mathrm{enc3}}\}$$8 and $$\{G_{\mathrm{enc1}}, G_{\mathrm{enc2}}, G_{\mathrm{enc3}}\}$$9, followed by attention-based refinement (Wang et al., 15 Jun 2025). The resulting refined feature is then used together with the coarse aggregated feature to generate a pixel-wise attention map in DPAM. In this design, global guidance and pixel-level guidance are explicitly separated.

CFSAM in (Xie et al., 16 Oct 2025) is less directional and more symmetric. It flattens multiple detection layers, concatenates them into one cross-layer token sequence, applies partitioned Transformer-based self-attention, and then restores per-scale outputs through concatenation and $$V_{\mathrm{agg}} = \mathrm{Concat}_c \left( \{ \mathrm{GAP}(F_{\mathrm{enc}i}) \}_{i=1}^{3} \right)$$0 projection. It is therefore most accurately described as implicit global multi-scale guidance via cross-layer self-attention rather than explicit top-down semantic guidance.

5. Common design principles and recurrent misconceptions

Across these works, a plausible cross-paper abstraction is that GMGM-like designs contain four recurrent stages: global descriptor extraction, scale alignment or cross-layer mixing, adaptive weighting or semantic propagation, and stage-specific reinjection into the downstream hierarchy. The global descriptor may come from global average pooling over shallow features, full self-attention over a deepest feature map, hierarchical multi-branch pooling, or cross-layer token concatenation (Sang et al., 16 Apr 2026). The multi-scale stage may involve explicit pyramid levels such as $$V_{\mathrm{agg}} = \mathrm{Concat}_c \left( \{ \mathrm{GAP}(F_{\mathrm{enc}i}) \}_{i=1}^{3} \right)$$1, shallow encoder scales, or directional multi-kernel feature branches (Tao et al., 4 Mar 2026).

A frequent misconception is that “global” necessarily implies transformer-style self-attention. The literature here does not support that narrow reading. In HaLoBuild-Net, globality is produced by global average pooling on three encoder scales and by a lightweight cross-scale attention network with two $$V_{\mathrm{agg}} = \mathrm{Concat}_c \left( \{ \mathrm{GAP}(F_{\mathrm{enc}i}) \}_{i=1}^{3} \right)$$2 convolutions and a ReLU (Sang et al., 16 Apr 2026). In FSGNet, it comes from hierarchical average pooling at multiple resolutions in the GPM. In MGDFIS, it comes from directional regrouping and convolution rather than from explicit pairwise affinity matrices. Self-attention is one implementation route, but not the only one.

A second misconception is that “multi-scale” always means a full feature pyramid network. In the supplied literature, multi-scale can mean three shallow encoder features, $$V_{\mathrm{agg}} = \mathrm{Concat}_c \left( \{ \mathrm{GAP}(F_{\mathrm{enc}i}) \}_{i=1}^{3} \right)$$3 cross-scale attention, a four-branch pooling pyramid, or a pairwise fusion block inserted repeatedly into a detector neck. The relevant constant is not the exact topology but the use of more than one semantic/spatial scale under an explicit guidance rule.

A third misconception is that “guidance” must be an interpretable scalar attention map. In these papers, guidance takes several forms: concatenated guidance tensors in decoder stages, semantic flows from a bottleneck source, query-conditioned sparse sampling, channel gating, and pixel-wise weighting maps. A plausible implication is that GMGM should be treated as a family of guidance strategies rather than a single canonical operator.

6. Empirical evidence, strengths, and limitations

The clearest quantitative evidence for the explicit GMGM appears in HaLoBuild-Net’s ablations (Sang et al., 16 Apr 2026). Adding GMGM alone to the baseline improves performance on both adverse-condition subsets. On HaLo-L, IoU rises from $$V_{\mathrm{agg}} = \mathrm{Concat}_c \left( \{ \mathrm{GAP}(F_{\mathrm{enc}i}) \}_{i=1}^{3} \right)$$4 to $$V_{\mathrm{agg}} = \mathrm{Concat}_c \left( \{ \mathrm{GAP}(F_{\mathrm{enc}i}) \}_{i=1}^{3} \right)$$5 and F1 rises from $$V_{\mathrm{agg}} = \mathrm{Concat}_c \left( \{ \mathrm{GAP}(F_{\mathrm{enc}i}) \}_{i=1}^{3} \right)$$6 to $$V_{\mathrm{agg}} = \mathrm{Concat}_c \left( \{ \mathrm{GAP}(F_{\mathrm{enc}i}) \}_{i=1}^{3} \right)$$7. On HaLo-H, IoU rises from $$V_{\mathrm{agg}} = \mathrm{Concat}_c \left( \{ \mathrm{GAP}(F_{\mathrm{enc}i}) \}_{i=1}^{3} \right)$$8 to $$V_{\mathrm{agg}} = \mathrm{Concat}_c \left( \{ \mathrm{GAP}(F_{\mathrm{enc}i}) \}_{i=1}^{3} \right)$$9 and F1 rises from $1\times1$0 to $1\times1$1. The paper interprets this as validation that “global topological anchoring is essential for compensating for the visibility loss and structural ambiguity in degraded imagery.”

The gains are larger when GMGM is combined with the other decoder-side modules. With SFFM + GMGM, IoU reaches $1\times1$2 on HaLo-L and $1\times1$3 on HaLo-H. With MGFM + GMGM, IoU reaches $1\times1$4 and $1\times1$5. With the full SFFM + MGFM + GMGM configuration, IoU reaches $1\times1$6 on HaLo-L and $1\times1$7 on HaLo-H (Sang et al., 16 Apr 2026). The module therefore contributes complementary information rather than duplicating restoration or mutual-guided fusion.

Related papers report analogous evidence for GMGM-like functionality. In (Tao et al., 4 Mar 2026), GRM and CSHA each improve small-object detection performance, with the best GRM insertion point occurring after SPPF on $1\times1$8, reinforcing the interpretation that global guidance works best when applied to semantically aggregated features. In (Zhang et al., 26 Mar 2026), increasing the number of Global Semantic Guidance Flows consistently improves detector performance, with four flows yielding the best results, which directly supports multi-scale semantic propagation. In (Wang et al., 15 Jun 2025), the addition of GMM, DMM, FTSSA, and DPAM increases AP from $1\times1$9 to $1\times1$0, indicating that global mixing, detail preservation, and pixel guidance are additive rather than interchangeable.

The limitations are equally consistent across the literature. In HaLoBuild-Net, the paper does not specify the exact channel dimensions of GMGM branches, the reduction ratio inside the $1\times1$1 weighting network, whether $1\times1$2 are scalar or channel-wise, or the exact downsampling operator $1\times1$3 (Sang et al., 16 Apr 2026). More generally, many papers present GMGM-like behavior only as a composite of multiple modules rather than as a monolithic block. This suggests that GMGM is presently better understood as an architectural motif—global context extraction followed by scale-aware guidance injection—than as a single fixed implementation.

In that sense, GMGM occupies a precise but evolving role in contemporary model design. In its explicit HaLoBuild-Net form, it is a topology-anchoring guidance branch for adverse-condition building extraction. In broader usage, it names a recurrent design principle: extract a global prior from one or more scales, transform it into scale-aware guidance, and reinject it into the feature hierarchy so that local decoding remains constrained by scene-level structure.