Local2Global: Integrating Local and Global

• Local2Global is a research paradigm that transforms local measurements into coherent global representations across multiple domains.
• It employs a stepwise approach: defining local units, processing them independently, and synchronizing results through explicit compatibility operations.
• Applications span scalable graph embedding, indoor panorama analysis, image retrieval, video segmentation, and even global forecasting in migration and solar irradiance.

Searching arXiv for recent and relevant "Local2Global" papers to ground the article. Local2Global denotes a recurring research pattern in which a system begins with local structure—overlapping graph patches, local descriptors, consecutive vertical panels, case-study sites, community-scale ties, robot-centric measurements, or boundary data on an open subset—and then constructs a global representation, global inference, or global coordinate frame. The term appears in graph representation learning (Jeub et al., 2021), indoor 360 panorama understanding (Yu et al., 2023), image retrieval (Aiger et al., 4 Sep 2025), video instance segmentation (Koner et al., 27 Jul 2025), land change science (Magliocca et al., 2013), migration-network analysis (Danchev et al., 2016), swarm robotics (Lentzas et al., 2021), and Dirichlet-to-Neumann theory (Daudé et al., 28 Jun 2026).

1. Scope and domain-specific meanings

The literature does not use Local2Global as the name of a single algorithm. Instead, it names a family of constructions in which local measurements or local models are made compatible with a global objective. In graph learning, the objective is a globally consistent embedding assembled from independently trained patch embeddings (Jeub et al., 2021). In PanelNet, the objective is panel-wise global context over 360 indoor panoramas (Yu et al., 2023). In image retrieval, the objective is to flip the dominant global-to-local pipeline into a local-to-global retrieval paradigm (Aiger et al., 4 Sep 2025). In video instance segmentation, the objective is early alignment between current-frame local queries and propagated global queries (Koner et al., 27 Jul 2025).

This breadth is visible outside machine learning as well. GLOBE formalizes the step from local insight to global inference by embedding case studies into a global analytic grid and computing representativeness indicators (Magliocca et al., 2013). World migration research uses Local2Global dynamics to describe heterogeneous interplay of local and global tendencies across “cave,” “biregional,” and “bridging” communities (Danchev et al., 2016). Swarm robotics uses a rigid-body transformation from robot-centric localization to a shared global frame (Lentzas et al., 2021). In inverse problems, local equality of Dirichlet-to-Neumann maps is propagated to global equality on the boundary (Daudé et al., 28 Jun 2026).

Domain	Local objects	Global target
Graph representation learning	overlapping subgraphs or “patches”	globally consistent representation
Indoor 360 vision	consecutive vertical panels	panel-wise global context
Image retrieval	local descriptors and local similarities	global feature re-ranking
Video instance segmentation	local queries and global queries	temporally consistent predictions
Land change science	local and regional observations	globally relevant integration
Swarm robotics	local frame $\{R\}$	common global frame $\{G\}$

2. Recurrent methodological structure

A common Local2Global pipeline first defines local units, then performs local processing, and finally aligns, aggregates, or propagates these local results into a global object. This suggests a methodological template rather than a single architecture. In graph embedding, the local units are overlapping patches; in image retrieval, they are local descriptor matches; in PanelNet, they are panels and within-panel windows; in video instance segmentation, they are current-frame local queries; in GLOBE, they are georeferenced case sites.

The global step varies, but it is usually an explicit compatibility operation. Graph methods estimate rigid motions from overlaps and then average aligned node coordinates. L2G2G writes the merged embedding of node $u$ as

$$\bar Z_u=\frac{1}{|\{\,i:u\in P_i\}|}\sum_{i:u\in P_i}\bigl(S_i\,Z^{(i)}_u+t_i\bigr),$$

where $S_i\in O(r)$ and $t_i\in\mathbb R^r$ are synchronized patch transforms (OuYang et al., 2024). Image retrieval constructs a dissimilarity matrix from local similarities and applies multidimensional scaling to obtain on-the-fly global features by minimizing

$$\min_{z_1,\dots,z_n\in\mathbb R^p}\sum_{i<j}(\|z_i-z_j\|_2-d_{ij})^2,$$

with SMACOF used in practice for non-metric and incomplete distances (Aiger et al., 4 Sep 2025).

The same pattern appears in transformer-based models. PanelNet first performs Window Blocks for local aggregation within each panel and then Panel Blocks for global aggregation across panels (Yu et al., 2023). The Local2Global framework for video instance segmentation aligns $Q_{\text{global}}$ with $Q_{\text{local}}$ in a lightweight transformer decoder before the main segmentation decoder, so that propagated queries are re-anchored to the current frame (Koner et al., 27 Jul 2025). In each case, the global stage is not merely a larger receptive field; it is an explicit operation that makes local information mutually compatible.

3. Patch alignment and scalable graph representation learning

In graph representation learning, Local2Global is a decentralized scaling strategy. “Local2Global: Scaling global representation learning on graphs via local training” divides a large graph into overlapping subgraphs, trains local representations for each patch independently, and combines them into a globally consistent representation by estimating rigid motions using information from patch overlaps via group synchronization (Jeub et al., 2021). A key distinguishing feature is that patches are trained independently without the need for parameter synchronization during distributed training.

The distributed formulation is developed further in “Local2Global: A distributed approach for scaling representation learning on graphs” (Jeub et al., 2022). There, the graph is covered by patches $P_1,\dots,P_p\subseteq V$, the minimum overlap condition $\{G\}$0 is imposed on the patch graph, local embeddings $\{G\}$1 are trained patch-wise, and global alignment is carried out by scale synchronization, orthogonal synchronization, and translation synchronization. The final global embedding is the centroid of aligned patch coordinates for each node. The same paper reports that on MAG240m, Local2Global markedly improves over the “no-align” baseline for both VGAE and DGI, and that on small datasets it is within $\{G\}$2–$\{G\}$3 of “full” training (Jeub et al., 2022).

L2G2G modifies this template by synchronizing during training rather than after training. “L2G2G: a Scalable Local-to-Global Network Embedding with Graph Autoencoders” aligns latent node representations in each epoch, uses a local patch loss, and reports higher accuracy than the standard Local2Global approach while maintaining scalability (OuYang et al., 2024). On SBM-Large-Dense, the reported link-prediction AUC is $\{G\}$4 for L2G2G versus $\{G\}$5 for GAE+L2G, and the method remains within a small constant factor of GAE+L2G in per-epoch runtime while being much faster than a full GAE on the largest networks (OuYang et al., 2024).

A broader conceptual perspective is given in “Local-to-global Perspectives on Graph Neural Networks” (Cai, 2023). That work distinguishes local MPNN from global Graph Transformer, studies convergence of Invariant Graph Networks, connects MPNN and Graph Transformer, and uses local MPNN for graph coarsening. One of its central statements is that MPNN+VN can approximate one layer of Performer/Linear Transformer to arbitrary accuracy, which situates Local2Global graph learning inside a wider theory of local and global graph computation (Cai, 2023).

4. Visual computing: panels, retrieval, and propagated queries

In visual computing, Local2Global frequently appears as a correction to local bias, spatial fragmentation, or temporal drift. PanelNet is a clear example. It represents an equirectangular panorama as consecutive vertical panels with corresponding 3D panel geometry, adds a panel geometry embedding network based on $\{G\}$6, and introduces a Local2Global Transformer that aggregates local information within a panel and panel-wise global context (Yu et al., 2023). The transformer stacks Window Blocks from fine to coarse resolutions $\{G\}$7 and then 6 Panel Blocks. Quantitatively, the full PanelNet reaches MRE $\{G\}$8 on Stanford2D3D, improves Matterport3D depth estimation from $\{G\}$9 for Omnifusion baseline to $u$0, attains $u$1 mIoU on semantic segmentation, and reports $u$2 3DIoU on layout estimation (Yu et al., 2023).

Image retrieval uses Local2Global in a more literal sense: it reverses the dominant retrieval order. “Global-to-Local or Local-to-Global? Enhancing Image Retrieval with Efficient Local Search and Effective Global Re-ranking” states that the dominant paradigm is global-to-local, whereas Local→Global first retrieves candidates by efficient local feature search and then re-ranks them with global features computed on-the-fly from local retrieval similarities (Aiger et al., 4 Sep 2025). The local similarity is the asymmetric Chamfer similarity

$u$3

and CANN gives $u$4 query time. The reported runtime is typically $u$5 on $u$6 images for local search and about $u$7 for on-the-fly MDS with $u$8; the paper reports new state-of-the-art results on Revisited Oxford and Paris and notes an improvement of about $u$9–$$\bar Z_u=\frac{1}{|\{\,i:u\in P_i\}|}\sum_{i:u\in P_i}\bigl(S_i\,Z^{(i)}_u+t_i\bigr),$$0 mAP over AMES in large-scale $$\bar Z_u=\frac{1}{|\{\,i:u\in P_i\}|}\sum_{i:u\in P_i}\bigl(S_i\,Z^{(i)}_u+t_i\bigr),$$1 settings (Aiger et al., 4 Sep 2025).

Video instance segmentation introduces a query-centric Local2Global variant. “Local2Global query Alignment for Video Instance Segmentation” defines $$\bar Z_u=\frac{1}{|\{\,i:u\in P_i\}|}\sum_{i:u\in P_i}\bigl(S_i\,Z^{(i)}_u+t_i\bigr),$$2 from the current frame, $$\bar Z_u=\frac{1}{|\{\,i:u\in P_i\}|}\sum_{i:u\in P_i}\bigl(S_i\,Z^{(i)}_u+t_i\bigr),$$3 from the previous frame’s propagated decoder queries, and an L2G-aligner consisting of a 3-layer transformer decoder with hidden dimension $$\bar Z_u=\frac{1}{|\{\,i:u\in P_i\}|}\sum_{i:u\in P_i}\bigl(S_i\,Z^{(i)}_u+t_i\bigr),$$4 by default (Koner et al., 27 Jul 2025). Dynamic positional embeddings are computed from $$\bar Z_u=\frac{1}{|\{\,i:u\in P_i\}|}\sum_{i:u\in P_i}\bigl(S_i\,Z^{(i)}_u+t_i\bigr),$$5 through $$\bar Z_u=\frac{1}{|\{\,i:u\in P_i\}|}\sum_{i:u\in P_i}\bigl(S_i\,Z^{(i)}_u+t_i\bigr),$$6. The method reports $$\bar Z_u=\frac{1}{|\{\,i:u\in P_i\}|}\sum_{i:u\in P_i}\bigl(S_i\,Z^{(i)}_u+t_i\bigr),$$7 AP on Youtube-VIS-2019, $$\bar Z_u=\frac{1}{|\{\,i:u\in P_i\}|}\sum_{i:u\in P_i}\bigl(S_i\,Z^{(i)}_u+t_i\bigr),$$8 AP on Youtube-VIS-2021, $$\bar Z_u=\frac{1}{|\{\,i:u\in P_i\}|}\sum_{i:u\in P_i}\bigl(S_i\,Z^{(i)}_u+t_i\bigr),$$9 AP on OVIS, and VPQ $S_i\in O(r)$0 on VIPSeg with a ResNet-50 backbone (Koner et al., 27 Jul 2025). Its main claim is that early alignment of local and global queries improves temporal consistency without additional complex heuristics or memory mechanisms.

A related, though differently named, design appears in “Global-Local Processing in Convolutional Neural Networks,” which proposes a Global Advantage Stream and a plug-and-play Global/Local Processing model to complement the local bias of CNNs (Rezvani et al., 2023). This suggests that Local2Global in vision is often tied to the same technical problem: recovering holistic structure that standard local processing under-emphasizes.

5. Spatial synthesis, migration structure, and generalized forecasting

Outside representation learning, Local2Global also names a strategy for making global claims from local evidence. In land change science, GLOBE uses a Discrete Global Grid of approximately $S_i\in O(r)$1 million equal-area hexagons of $S_i\in O(r)$2 each, integrates more than $S_i\in O(r)$3 publicly available global rasters, overlays local case sites on the grid, and computes sample and population distributions over selected global variables (Magliocca et al., 2013). Its representativeness analysis uses a Hellinger-distance-based indicator with

$S_i\in O(r)$4

and also defines representedness scores $S_i\in O(r)$5 for geovisualization of over- and under-sampling (Magliocca et al., 2013). The associated weighting export uses inverse-probability weights $S_i\in O(r)$6. In this context, Local2Global is a workflow for statistically robust, globally relevant integration of local and regional observations.

Migration-network analysis uses the term in a structural rather than algorithmic sense. The World Migration Network is a directed weighted multilayer network of $S_i\in O(r)$7 countries across the decades $S_i\in O(r)$8, with communities detected by multilayer modularity maximization and then classified through strength-based and overlap-based E–I indices (Danchev et al., 2016). For a community $S_i\in O(r)$9,

$t_i\in\mathbb R^r$0

Hierarchical clustering in the $t_i\in\mathbb R^r$1 plane yields cave communities, biregional communities, and bridging communities. Cave communities are tightly-knit and enduring, biregional communities connect two distinct geographic regions, and bridging communities have hub-and-spoke structures and high strength inequality (Danchev et al., 2016). The paper’s conclusion is that world migration is a patchwork of cave regions, bridging hubs, and biregional glocal transitions rather than a uniformly globalized system.

Generalized solar irradiance forecasting applies the local/global distinction to model training regimes. “Local-Global Methods for Generalised Solar Irradiance Forecasting” compares Local models, a single Global model, cv-mode with held-out sites, and kn-mode using nearest-neighbour irradiance (Cargan et al., 2023). Using data from $t_i\in\mathbb R^r$2 UK locations, a 6-hour forecast horizon, and both weather and satellite inputs, the Global CNN is reported as the best overall model with nRMSE $t_i\in\mathbb R^r$3 and Skill $t_i\in\mathbb R^r$4, while cv-mode performs almost identically to the Global case, indicating zero-shot generalisation to unseen sites (Cargan et al., 2023). The same study states that a single Global CNN or LSTM can be deployed to produce forecasts at a new solar farm immediately, without any historical irradiance.

6. Shared coordinate systems, propagation principles, and conceptual distinctions

In swarm robotics, Local2Global has a literal geometric meaning. “From Robot Self-Localization to Global-Localization: An RSSI Based Approach” begins from a robot-centric frame $t_i\in\mathbb R^r$5, uses four perimeter receivers and one center transceiver, estimates beacon positions by trilateration, and defines a global frame $t_i\in\mathbb R^r$6 using three stationary robots acting as beacons (Lentzas et al., 2021). The transformation from local to global coordinates is written as

$t_i\in\mathbb R^r$7

The evaluation uses Euclidean error in global position, reports exact recovery at $t_i\in\mathbb R^r$8 noise, error $t_i\in\mathbb R^r$9 at $$\min_{z_1,\dots,z_n\in\mathbb R^p}\sum_{i<j}(\|z_i-z_j\|_2-d_{ij})^2,$$0 noise for a robot static at $$\min_{z_1,\dots,z_n\in\mathbb R^p}\sum_{i<j}(\|z_i-z_j\|_2-d_{ij})^2,$$1 from all beacons, and sub-metre accuracy up to $$\min_{z_1,\dots,z_n\in\mathbb R^p}\sum_{i<j}(\|z_i-z_j\|_2-d_{ij})^2,$$2 under moderate noise (Lentzas et al., 2021).

In inverse problems, Local2Global denotes propagation of operator equality. “A Local–to–Global Propagation Principle for Dirichlet–to–Neumann Maps” proves three local-to-global propagation results: one under a collar hypothesis, one under an exponential spectral assumption, and one for conformally warped product metrics under quasi-analytic boundary closeness (Daudé et al., 28 Jun 2026). In the first theorem, if two smooth Riemannian metrics coincide in a collar neighborhood of a connected boundary component $$\min_{z_1,\dots,z_n\in\mathbb R^p}\sum_{i<j}(\|z_i-z_j\|_2-d_{ij})^2,$$3, then equality of local Dirichlet-to-Neumann maps on a nonempty open subset $$\min_{z_1,\dots,z_n\in\mathbb R^p}\sum_{i<j}(\|z_i-z_j\|_2-d_{ij})^2,$$4 propagates to equality of global Dirichlet-to-Neumann maps on all of $$\min_{z_1,\dots,z_n\in\mathbb R^p}\sum_{i<j}(\|z_i-z_j\|_2-d_{ij})^2,$$5 (Daudé et al., 28 Jun 2026). The second theorem replaces the collar hypothesis by the bound

$$\min_{z_1,\dots,z_n\in\mathbb R^p}\sum_{i<j}(\|z_i-z_j\|_2-d_{ij})^2,$$6

and the third specializes to warped-product metrics where Weyl–Titchmarsh theory and quasi-analytic propagation are used (Daudé et al., 28 Jun 2026).

These usages clarify several common misconceptions. Local2Global is not synonymous with hierarchical pooling, because in graphs it is rigid-motion synchronization, in retrieval it is search followed by MDS-based re-ranking, in robotics it is frame transformation, and in Dirichlet-to-Neumann theory it is a propagation theorem. It is not necessarily global-first; the image-retrieval work explicitly frames its contribution as a reversal of the dominant global-to-local paradigm (Aiger et al., 4 Sep 2025). Nor does it require monolithic global computation: the graph literature emphasizes that local models can be trained independently and only later aligned (Jeub et al., 2021). A plausible implication is that the term is best understood as a design principle for reconciling locality with system-level consistency, rather than as a fixed algorithmic family.