Global-Local Collaborative Feature Fusion (CFF)

• Global-Local Collaborative Feature Fusion is a design strategy that separates coarse global context from fine local details and fuses them to create richer task-specific embeddings.
• It employs diverse fusion topologies—including dual-stream encoder, gated fusion, and cross-layer fusion—to reconcile whole-sample structure with part-level specificity.
• CFF has demonstrated improved performance in applications such as autonomous driving and medical segmentation while balancing computational efficiency and model accuracy.

Searching arXiv for the requested paper and closely related recent work on global–local feature fusion. Searching arXiv for “Context-Aware Feature-Fusion for Co-occurring Object Detection in Autonomous Driving” and “Global-Local Collaborative Feature Fusion”. Global-Local Collaborative Feature Fusion (CFF) denotes a family of representation-learning strategies in which a model explicitly separates global context from local detail and then learns a joint representation from both. Across recent work, the “global” component usually summarizes scene-level, trial-level, or graph-level structure, while the “local” component preserves part-level, region-level, channel-level, or neighborhood-level specificity. The fused representation is then used as the task-facing embedding for detection, segmentation, recognition, estimation, or recommendation. In autonomous driving, for example, “Context-Centric Feature Fusion (CCFF)” combines a Local Context Fusion Module and a Global Context Attention Module for co-occurring object detection (Singh et al., 10 Jun 2026).

1. CFF as a recurring architectural pattern

CFF is not a single architecture but a recurring design pattern. In bone age assessment, BoNet+ uses “global and local feature extraction channels” and fuses them before regression (Lou et al., 20 Dec 2025). In medical segmentation, DyGLNet uses a “hybrid feature extraction module” that splits channels into global and local branches inside each block (Zhao et al., 16 Sep 2025). In camera-only 3D detection, Collaborative Perceiver defines a “global-local collaborative feature fusion (CFF) module” that integrates height-collapsed BEV features with height-aware local BEV features (Yuan et al., 28 Jul 2025). In hyperspectral classification, “cross-layer feature fusion (CFF)” refers to adaptive fusion of transformer inputs and intermediate outputs so that shallow and deep features are jointly preserved (Chen et al., 26 Apr 2026).

A concise way to view the literature is that CFF introduces an explicit inductive bias: coarse context and fine structure are treated as complementary rather than interchangeable. This suggests that CFF is best understood as a representational decomposition followed by learned recomposition, rather than merely as a late concatenation trick.

Domain	Global component	Local component
Autonomous driving (Singh et al., 10 Jun 2026)	GCAM pools top-K RoI features into a global context attention token	LCFM uses RoI-to-RoI self-attention
Bone age assessment (Lou et al., 20 Dec 2025)	Transformer global feature extraction channel	RFAConv local feature extraction channel
Medical segmentation (Zhao et al., 16 Sep 2025)	single-head self-attention branch	multi-scale dilated depthwise convolutions branch
Vision-based 3D detection (Yuan et al., 28 Jul 2025)	height-collapsed BEV features $f_g$	VHS-derived local BEV features $f_l$

2. Canonical fusion topologies and operators

A common topology is the dual-stream encoder. BoNet+ processes a cropped radiograph $I$ with a global stream and an attention-map input $A$ with a local stream, then fuses them by channel concatenation,

$F_{\text{fusion}} = \text{Concat}(F_g, F_l),$

before joint refinement by Inception-V3 (Lou et al., 20 Dec 2025). DyGLNet internalizes the same principle at block level: channels are split into a global branch with single-head self-attention and a local branch with three dilated depthwise convolutions, then recombined by

$$X_{\mathrm{fused}} = \mathrm{Conv}_{1\times1}\left([\mathrm{Attention}(Q, K, V); X_l^{\text{out}}]\right).$$

This yields global-local fusion inside the encoder rather than only at the head (Zhao et al., 16 Sep 2025).

A second topology is gated fusion. Collaborative Perceiver refines global and local BEV features separately, computes

$$\alpha=\sigma\left(f^{con}_l \oplus f^{con}_g\right),$$

and fuses them by

$$f_{u}^{bev}=\alpha \odot Conv\left(f_{g}\right)+\left(1-\alpha\right) \odot Conv\left(f_{l}\right).$$

Here fusion is neither pure concatenation nor pure addition; it is a learned per-element selection between complementary sources (Yuan et al., 28 Jul 2025). A related quality-aware formulation appears in LGAF for face recognition, where local and global feature norms are normalized into weights $\gamma_i^l,\gamma_i^g$ and the final embedding is

$$\kappa_i = \gamma_i^l \Psi_i + \gamma_i^g \Upsilon_i.$$

This operationalizes fusion as adaptive reweighting conditioned on feature quality (Yu et al., 2024).

A third topology is cross-layer or cross-view fusion. In the hyperspectral model, the transformer input $f_l$0 and intermediate encoder outputs $f_l$1 are concatenated and linearly fused into $f_l$2 before the last encoder. The purpose is to reduce information loss across depth rather than to combine different spatial resolutions or modalities (Chen et al., 26 Apr 2026). In federated learning, an analogous principle appears when local representations $f_l$3 are aligned by per-client matrices $f_l$4, then fused through a consensus graph and GCN, so that local embeddings become globally comparable before aggregation (Ma et al., 2023).

These examples show that CFF spans at least three operator families: concatenation-based fusion, gated fusion, and structure-aware fusion over token, graph, or layer relations. It is therefore misleading to equate CFF exclusively with attention, even though attention is common.

3. Operational meanings of “global” and “local”

The meanings of “global” and “local” vary by task, but the distinction is systematic. In CCFF for autonomous driving, the local component is object-centric: the Local Context Fusion Module “uses the RoI-to-RoI self-attention mechanism to resolve spatial interactions,” especially for “small and partially obscured objects,” while the Global Context Attention Module pools top-$f_l$5 RoI features into a “global context attention token” to capture co-occurrence priors without pixel-level global pooling (Singh et al., 10 Jun 2026). In this setting, locality is relational and RoI-based rather than pixel-based.

In EEG emotion recognition, locality is channel-wise. The local branch is a $f_l$6 descriptor built from differential entropy and node-level graph-theoretic features, flattened into a 558-dimensional EEG vector, whereas the global branch is a 25-dimensional trial-level descriptor formed by average pooling time-domain, spectral, and multifractal features across channels (Zhou et al., 13 Jan 2026). Here globality means cross-channel statistical summarization, not large receptive fields.

In BEV 3D detection, globality is tied to height collapse and scene layout, while locality is tied to height-aware occupancy structure. Collaborative Perceiver explicitly defines $f_l$7 as height-collapsed BEV context and $f_l$8 as local BEV features extracted by Voxel-Height-guided Sampling from height intervals of interest (Yuan et al., 28 Jul 2025). In mining-scene classification, the global branch is a Swin-based multi-scale transformer with a collaborative dictionary of key semantic vectors, whereas the local branch uses CNN features reweighted by their contribution to the global semantic basis (Fan et al., 27 Jul 2025).

A useful synthesis is that “global” usually denotes a representation whose support spans the full sample or graph, while “local” denotes a representation indexed by parts, regions, channels, or proposals. This suggests that CFF is fundamentally about reconciling support mismatch: whole-sample structure and fine-grained evidence are encoded at different granularities and must be made jointly usable.

4. Learning objectives and optimization regimes

Most CFF systems keep the task loss of the host model and introduce fusion by architectural design rather than by a standalone fusion loss. BoNet+ uses a single regression target after fusion, optimized with Smooth L1 loss,

$f_l$9

so gradients from the fused prediction head train both branches collaboratively (Lou et al., 20 Dec 2025). DyGLNet similarly uses a hybrid segmentation objective,

$I$0

with no separate branch-specific supervision in the fusion block (Zhao et al., 16 Sep 2025).

Other systems make the collaborative nature explicit in the training objective. Collaborative Perceiver jointly optimizes detection and occupancy,

$I$1

so the fused BEV representation is shaped simultaneously by box-level and voxel-level supervision (Yuan et al., 28 Jul 2025). The EEG framework adds domain-adversarial regularization through

$I$2

with a gradient reversal layer so that the fused representation becomes more subject-invariant (Zhou et al., 13 Jan 2026). In mining-scene classification, the total objective is reported as

$I$3

with separate branch losses and a fusion-head loss, reflecting a stronger multi-loss interpretation of collaboration (Fan et al., 27 Jul 2025).

These training patterns imply two distinct regimes. In the first, CFF is an architectural prior trained by the host task loss alone. In the second, CFF is part of a multi-objective system in which complementary branches or tasks are kept individually discriminative while also being fused.

5. Empirical behavior, benefits, and efficiency trade-offs

Across domains, the most consistent empirical observation is that local-only and global-only models are both weaker than collaborative fusion. In BoNet+, adding only the Transformer branch yields MAE $I$4 months on RSNA and $I$5 months on RHPE, adding only RFAConv yields $I$6 and $I$7, and the full fused model reaches $I$8 and $I$9 months, respectively (Lou et al., 20 Dec 2025). In DyGLNet, removing self-attention or removing dilated convolutions degrades Dice across datasets, and replacing DyFusionUp with bilinear or transposed upsampling also lowers performance, indicating that global modeling, local modeling, and detail-preserving reconstruction are complementary rather than redundant (Zhao et al., 16 Sep 2025).

Autonomous-driving detection provides an explicitly object-centric example. CCFF reports “significant improvement on relational consistency,” with Category-level Consistency Strategy values of $A$0 on Cityscapes and $A$1 on BDD100K, “substantial gains in small object detection (AP_S: 14.1\%),” recovery of rare classes such as “Train,” and real-time processing with a “0.2 FPS overhead” (Singh et al., 10 Jun 2026). In camera-only 3D detection, adding CFF on top of dense occupancy, LDO, and VHS increases validation performance from $A$2 to $A$3 NDS and from $A$4 to $A$5 mAP; the full CoP system reaches $A$6 mAP and $A$7 NDS on the nuScenes test set (Yuan et al., 28 Jul 2025). In face recognition, LGAF achieves the best average performance on CFP-FP, CPLFW, AgeDB, and CALFW, and reaches $A$8 Rank-1 on TinyFace in the ArcFace setting, supporting the claim that quality-aware global-local fusion is especially useful for low-quality faces (Yu et al., 2024).

At the same time, CFF is not free. DyGLNet notes that dynamic upsampling adds overhead, even though the overall design remains lightweight at $A$9M parameters and $F_{\text{fusion}} = \text{Concat}(F_g, F_l),$0G FLOPs (Zhao et al., 16 Sep 2025). BoNet+ explicitly notes that “two CNN branches plus a Transformer and Inception‑V3 is heavier than single-stream baselines” (Lou et al., 20 Dec 2025). CCFF’s object-centric design partly addresses this trade-off by replacing pixel-level global pooling with top-$F_{\text{fusion}} = \text{Concat}(F_g, F_l),$1 RoI pooling (Singh et al., 10 Jun 2026). A plausible implication is that successful CFF designs usually win by allocating computation selectively: to salient RoIs, informative height ranges, branch-specific tokens, or sparse graphs, rather than by uniformly increasing model capacity.

6. Misconceptions, limitations, and open directions

A common misconception is that CFF is synonymous with simple concatenation. The literature repeatedly contradicts this. The hyperspectral model introduces cross-layer CFF precisely because “relying solely on simple concatenation may weaken the connectivity between these layers” (Chen et al., 26 Apr 2026). GraphTransfer argues that concatenation and summation are “too simple to fully capture the non-linearity between different types of features,” and instead aligns interaction scores across feature spaces (Xia et al., 2024). LGAF further shows that equal addition of local and global features is weaker than quality-weighted fusion (Yu et al., 2024).

Another misconception is that the “global” branch is always more robust. The evidence is conditional. In face recognition, missing local facial regions favor local similarity, while local deformation can make global features more reliable (Yu et al., 2024). In EEG, local channel-wise descriptors are rich but sensitive to subject-specific variability, whereas global descriptors are more stable but less fine-grained (Zhou et al., 13 Jan 2026). Collectively, these papers suggest that the central question is not whether global beats local, but when one should dominate the fusion.

Limitations also recur. BoNet+ reports degradation in underrepresented age groups and attention leakage to irrelevant background (Lou et al., 20 Dec 2025). DyGLNet still struggles on very low-contrast images (Zhao et al., 16 Sep 2025). LGAF notes that fused local-global features can still be corrupted by unidentifiable or mislabeled samples (Yu et al., 2024). In federated feature fusion, scalability can become problematic because graph learning is naively $F_{\text{fusion}} = \text{Concat}(F_g, F_l),$2 in the number of clients (Ma et al., 2023). These are not isolated failures; they indicate that CFF inherits the weaknesses of both representation branches and can amplify them if alignment, gating, or supervision is poor.

Open directions in the cited work are comparatively consistent. BoNet+ suggests “cross-attention between global and local features” and “dynamic gating where global features select which local details to emphasize” (Lou et al., 20 Dec 2025). DyGLNet suggests further optimization of dynamic upsampling and stronger handling of low-contrast inputs (Zhao et al., 16 Sep 2025). The EEG framework points toward richer spatiotemporal or graph-aware transformers for preserving electrode topology (Zhou et al., 13 Jan 2026). More generally, this suggests that the next stage of CFF research will likely emphasize explicit alignment mechanisms, finer-grained gating, and multi-task or multi-modal supervision, rather than larger undifferentiated dual-branch models.