Dual-Granularity Identity Association

Updated 6 July 2026

The paper introduces a weakly supervised text-to-person matching framework that leverages dual-scale (local and global) identity inference to overcome brittleness of clustering methods.
Local associations build explicit cross-modal identity relations within mini-batches using contrastive learning and similarity thresholding on image and text features.
Global associations use memory banks and dynamic confidence adjustment to propagate one-to-many identity cues, achieving up to a 3% improvement in Rank-1 accuracy.

Searching arXiv for the named mechanism and closely related papers. I’ll look up arXiv records relevant to “local-and-global dual-granularity identity association.” The Local-and-Global Dual-Granularity Identity Association Mechanism most directly denotes a weakly supervised text-to-person image matching framework that infers identity relations at two coupled scales: a local level, where cross-modal relations are explicitly constructed within a mini-batch, and a global level, where broader one-to-many associations are mined across the training set through a memory bank and a dynamic confidence adjustment mechanism. In this formulation, the central problem is that only paired image-text samples are known during training, while most cross-modal identity relations among different samples are unknown; the mechanism is therefore designed to recover both subtle local correspondences and dataset-level identity structure without full identity supervision (Zhang et al., 9 Jul 2025). Related literature uses closely analogous local/global or fine/coarse association designs in video person re-identification, visible-thermal re-identification, image-language supervision, identity customization, long-term memory retrieval, and cross-document event coreference, although the operational meaning of “local” and “global” varies by task (Li et al., 2021).

1. Problem formulation and conceptual scope

In the weakly supervised text-to-person setting, the training set is

$\mathcal{D} = { (I_i, T_i) }_{i=1}^N,$

where $I_i$ is the $i$ -th person image and $T_i$ is its paired text description. The only guaranteed identity relation is the paired one: $(I_i, T_i)$ share the same pedestrian identity, whereas for $i \neq j$ , the identity relation between $I_i$ and $T_j$ is unknown. The mechanism is motivated by the fact that one identity can correspond to multiple images and multiple descriptions across cameras and wording variations, but weak supervision does not provide those extra labels (Zhang et al., 9 Jul 2025).

The local-and-global construction is intended to remedy a limitation of existing weakly supervised methods that usually rely on dataset-wide clustering to mine pseudo identity associations. The cited formulation states that such clustering is brittle for hard samples—including visually similar people or semantically ambiguous descriptions—because clustering can merge wrong identities or miss true ones. The proposed response is not a single pseudo-labeling step, but a two-scale association design comprising Basic Feature Extraction (BFE) with local relationship construction, Global Sample Relationship Construction and Feature Refinement (GSRC-FR), and Information-Asymmetric Sample Pair Construction and Consistency Learning (IASC-CL) (Zhang et al., 9 Jul 2025).

A broader conceptual reading is supported by adjacent work. In video person re-identification, local alignment and global appearance are described as complementary because local-only methods can be fragile under shifted or partial detections, while global-only methods may correlate background clutter or switched identities (Li et al., 2021). In long-term conversational agents, single-granularity retrieval is likewise described as too brittle because it loses either fine detail or high-level context (Xu et al., 26 May 2025). These neighboring formulations suggest that a dual-granularity mechanism is best understood not as a domain-specific heuristic, but as a structured response to the general failure mode of single-scale identity or relevance estimation.

2. Local-granularity association within the mini-batch

The local mechanism in the weakly supervised text-to-person formulation operates within a mini-batch and explicitly constructs cross-modal identity relations from three ingredients: intra-modal similarity among images, intra-modal similarity among texts, and known paired image-text supervision. The image encoder and text encoder are CLIP-pretrained, producing image features $\bm{f}_i^v \in \mathbb{R}^{1 \times d}$ and text features $\bm{f}_i^t \in \mathbb{R}^{1 \times d}$ . Basic alignment is first established through contrastive learning:

$I_i$ 0

with a symmetric $I_i$ 1, yielding

$I_i$ 2

This baseline, however, still uses paired samples directly and does not by itself recover latent one-to-many identity relations (Zhang et al., 9 Jul 2025).

The local association stage then computes intra-modal similarity matrices:

$I_i$ 3

Because self-similarity may contain noise from non-identity factors, the matrices are binarized by thresholding with $I_i$ 4:

$I_i$ 5

for $I_i$ 6. The cross-modal identity correspondence matrix is then defined by Hadamard product:

$I_i$ 7

This construction encodes the paper’s core local claim: if two samples are judged similar both in image space and text space, they are more likely to share identity (Zhang et al., 9 Jul 2025).

To reduce mismatches, the local relationship matrix is softened:

$I_i$ 8

after which zero entries are masked out,

$I_i$ 9

and converted into a target distribution

$i$ 0

The model prediction over text-image pairs is

$i$ 1

The local similarity distribution matching loss is

$i$ 2

with a symmetric $i$ 3, giving

$i$ 4

The stated effect is to form bidirectional cross-modal identity constraints inside a batch while also using within-modality structure to infer additional likely identity relations (Zhang et al., 9 Jul 2025).

The global mechanism addresses a structural limitation of the batch-level design: a mini-batch is too small to contain many same-identity cross-modal pairs. To expand supervision across the entire training set, the method introduces memory banks

$i$ 5

and updates the image memory bank by momentum:

$i$ 6

A key design choice is to use person images as the anchor, because visual appearance is described as carrying richer and more identity-specific cues than text (Zhang et al., 9 Jul 2025).

For a current batch image feature $i$ 7, similarity to all memory-bank image features is computed as

$i$ 8

For each row, the top- $i$ 9 nearest candidates are taken and then filtered by thresholding:

$T_i$ 0

These $T_i$ 1 are treated as samples sharing identity with image $T_i$ 2. Identity is then propagated from image $T_i$ 3 to its corresponding text, and from those texts back to other images, thereby establishing a dataset-level one-to-many identity association (Zhang et al., 9 Jul 2025).

Cross-modal association is expanded by combining within-batch text-image similarity

$T_i$ 4

with similarity between batch texts and memory-bank-selected images

$T_i$ 5

These are concatenated:

$T_i$ 6

and used to build a target matrix

$T_i$ 7

for each text’s candidate set $T_i$ 8 (Zhang et al., 9 Jul 2025).

The defining refinement is the confidence-based dynamic adjustment mechanism. For $T_i$ 9 and $(I_i, T_i)$ 0,

$(I_i, T_i)$ 1

The accompanying interpretation is explicit: directly using $(I_i, T_i)$ 2 may overemphasize high-confidence pairs and ignore weak but informative ones, whereas the adjusted target controls very strongly similar candidates and prevents weaker related pairs from being discarded completely. The final normalized target and prediction are

$(I_i, T_i)$ 3

and the global SDM loss is

$(I_i, T_i)$ 4

with symmetric $(I_i, T_i)$ 5, giving

$(I_i, T_i)$ 6

The stated role of this global stage is to expand supervision beyond the batch and improve sensitivity to weakly associated samples (Zhang et al., 9 Jul 2025).

4. Information asymmetry, consistency learning, and the complete training objective

The cited mechanism does not stop at local and global relation construction. It adds Information-Asymmetric Sample Pair Construction and Consistency Learning (IASC-CL) to address hard sample mining and robustness. For a paired sample $(I_i, T_i)$ 7, image augmentation is

$(I_i, T_i)$ 8

including horizontal flipping, border padding, random cropping, and random erasing, while text augmentation is

$(I_i, T_i)$ 9

which masks parts of the sentence with [MASK]. The asymmetric pair

$i \neq j$ 0

therefore has weaker explicit semantic correspondence while preserving identity consistency (Zhang et al., 9 Jul 2025).

The purpose of this construction is stated in practical terms. It simulates real-world conditions such as occlusion, viewpoint changes, incomplete descriptions, and varying description granularity, without requiring manually mined hard pairs. The consistency constraint on these asymmetric pairs is written as

$i \neq j$ 1

and is intended to preserve semantic identity consistency even when the available information differs across modalities (Zhang et al., 9 Jul 2025).

The full objective combines the baseline contrastive term, the local within-batch relation construction term, the global relation construction and refinement term, and the asymmetric consistency term:

$i \neq j$ 2

In the cited presentation, the improvement mechanism is cumulative. The local level captures fine-grained, immediate identity constraints inside the batch; the global level expands association beyond the batch and identifies broader one-to-many correspondences; the confidence-based adjustment keeps weak associations active; and the information-asymmetric learning supplies hard conditions without relying on noisy clustering (Zhang et al., 9 Jul 2025). This suggests that the “mechanism” is not merely the coexistence of two scales, but the interaction of local inference, global expansion, and robustness-oriented refinement.

5. Empirical profile and ablation evidence

The mechanism is evaluated on CUHK-PEDES, ICFG-PEDES, and RSTPReid. On CUHK-PEDES, the reported weakly supervised results are Rank-1: 73.06%, mAP: 64.88%, and mINP: 48.15%, improving on CPCL by Rank-1 by 3.03% and mAP by 1.69%. On ICFG-PEDES, the reported values are Rank-1: 63.71%, mAP: 36.87%, and mINP: 6.20%. On RSTPReid, the reported values are Rank-1: 61.30%, mAP: 47.20%, and mINP: 25.47%, improving on CPCL by Rank-1 by 2.95% and mAP by 1.39% (Zhang et al., 9 Jul 2025).

The ablation sequence is especially central because it isolates the contribution of each component. The baseline gives 64.69 / 54.24 / 54.60 Rank-1 on CUHK / ICFG / RSTPReid. Adding LRC yields 69.04 / 58.45 / 56.30. Adding GSRC-FR alone yields 67.56 / 58.51 / 55.90. Combining LRC + GSRC-FR yields 70.20 / 60.30 / 57.25. The full system with LRC + GSRC-FR + IASC-CL yields 73.06 / 63.71 / 61.30. The reported pattern is therefore additive rather than redundant: local association helps, global association helps, and the asymmetric consistency module provides the largest final boost (Zhang et al., 9 Jul 2025).

Additional analysis supports the reliability of the global mining stage. The paper reports that the global relation construction module achieves mined-relation accuracy of up to 92.0% on CUHK-PEDES and about 75.0% on ICFG-PEDES (Zhang et al., 9 Jul 2025). A plausible implication is that the memory-bank expansion is not functioning merely as a high-recall neighbor collector, but as a relatively precise identity propagation mechanism under weak supervision.

The phrase “Local-and-Global Dual-Granularity Identity Association Mechanism” is exact for weakly supervised text-to-person image matching, but several adjacent papers instantiate closely related structures.

Domain and paper	Local granularity	Global granularity
Weakly supervised text-to-person matching (Zhang et al., 9 Jul 2025)	Within-batch cross-modal identity relations	Memory-bank-expanded one-to-many relations
Video person Re-ID (Li et al., 2021)	Local part alignments via LAQ	Global appearance correlations via GCQ
VT-ReID with DGTL (Liu et al., 2020)	Sample-based triplet supervision	Center-based triplet supervision
Person Re-ID with language (Chen et al., 2018)	Image-region and noun-phrase association	Whole-image and whole-description association
VT-ReID with CM-LSP-GE (Wang et al., 2022)	Horizontal-part shortest-path alignment	BN-enhanced whole-person representation
Identity customization (Sun et al., 2024)	Face-region identity, pose, and expression control	Whole-image harmony and location control
Long-term memory retrieval (Xu et al., 26 May 2025)	Turn/session-level memory	Summary/keyword-level memory

In LOGA for video Re-ID, the method concurrently optimizes a local aligned quality (LAQ) module and a global correlated quality (GCQ) module, connected through a local-assembled global appearance prototype. The cited distinction from prior joint methods is that interaction matters more than mere parallelism: local evidence defines a clean anchor for the global branch, while global evidence feeds back to refine the final tracklet representation (Li et al., 2021). In DGTL for visible-thermal person re-identification, the dual-granularity structure is explicitly fine-to-coarse, combining sample-based triplet loss with center-based triplet loss while using only global person features; the same paper states that the method can still be viewed as associating identities at two granularity levels, moving from instance-level matching to center-level identity matching (Liu et al., 2020).

In image-language supervised person Re-ID, the mechanism is explicitly formulated as global and local image-language associations. The global image-language association is established according to the identity labels, while the local association aligns image regions with noun phrases and reconstructs phrases from phrase-specific visual features (Chen et al., 2018). In visible-thermal Re-ID with CM-LSP-GE, local identity association is implemented by horizontally partitioning the image and matching body-part graphs via a shortest-path strategy, whereas the global branch strengthens modality-shared identity discrimination with batch-normalization enhancement (Wang et al., 2022).

Outside re-identification, RealisID divides identity customization into a local branch for facial details, pose, expression, and scale robustness, and a global branch for image harmony and face location (Sun et al., 2024). MemGAS represents memory as a local-to-global bundle of session, turn, summary, and keywords, then uses GMM-based association, entropy-based granularity routing, Personalized PageRank, and LLM-based filtering to preserve local detail while integrating global conversational context (Xu et al., 26 May 2025). In event coreference, the relation between identity and granularity is framed through local mention evidence and global document-level aggregation, with the claim that event identity depends on compatibility of action, time, place, and participants rather than lexical matching alone (Vossen et al., 2017). Across these formulations, the recurring pattern is that fine-grained evidence supplies precision, while coarser aggregation supplies semantic completeness or robustness.

7. Conceptual distinctions and recurrent misconceptions

A recurrent misconception is to equate dual granularity with any architecture that has two branches. The literature does not support that simplification. In video Re-ID, the proposed associative strategy is explicitly reported to outperform a dual-branch strategy that learns local and global information separately, and the conclusion drawn there is that interaction matters more than just parallelism (Li et al., 2021). The weakly supervised text-to-person mechanism follows the same logic: the local and global stages are not independent heads attached to a common encoder, but successive relation-construction procedures whose outputs alter the supervision structure seen by the model (Zhang et al., 9 Jul 2025).

A second misconception is that “local” must mean body parts and “global” must mean holistic appearance. Some papers adopt that meaning, but the term is task-dependent. In the weakly supervised text-to-person setting, local refers to within-batch identity inference and global refers to dataset-level association through memory banks (Zhang et al., 9 Jul 2025). In DGTL, the relevant distinction is sample-based versus center-based supervision, and the paper explicitly notes that the method works with only global person features (Liu et al., 2020). In MemGAS, local and global denote turn/session details versus summary/keyword abstractions (Xu et al., 26 May 2025). The commonality is therefore not a fixed spatial decomposition, but the coordinated use of fine and coarse evidence.

A third misconception is that such mechanisms are simply alternatives to clustering. The weakly supervised text-to-person paper argues the opposite: dataset-wide clustering is brittle for hard samples, and the proposed mechanism is intended to replace sole reliance on clustering with local relationship construction, global memory-bank expansion, dynamic confidence adjustment, and information-asymmetric consistency learning (Zhang et al., 9 Jul 2025). Related work in event coreference reaches a comparable conclusion from a different angle: if identity constraints are too loose, the system produces “extreme aggregation and lumping,” whereas overly strict constraints make “each mention ... unique” (Vossen et al., 2017). In both cases, the central issue is controlled aggregation rather than indiscriminate merging.

Taken together, these papers define a dual-granularity identity association mechanism as a structured method for linking entities, samples, events, or memories across at least two representational scales. The exact implementation differs by domain, but the underlying design principle remains stable: use a finer level to preserve discriminative detail, use a coarser level to recover broader identity structure, and couple the two so that neither scale dominates the final association decision.