Multi-Center Contrastive Learning (MCCL)

• Multi-Center Contrastive Learning (MCCL) is a hierarchical contrastive mechanism that replaces single-center representations with multiple fine-grained prototypes to capture intra-class variations.
• MCCL employs a two-stage clustering process—using DBSCAN for coarse clustering and K-means for fine-grained sub-clustering—to enhance intra-modal supervision and cross-modal alignment.
• By integrating with BRST and a dual-path transformer, MCCL improves cross-modal matching and achieves notable performance gains on benchmarks like SYSU-MM01 and RegDB.

Searching arXiv for the specified paper and related MCCL terminology. {"query":"arXiv (Shi et al., 15 Sep 2025) Hierarchical Identity Learning for Unsupervised Visible-Infrared Person Re-Identification Multi-Center Contrastive Learning", "max_results": 5} {"query":"\"Multi-Center Contrastive Learning\" arXiv", "max_results": 10} Multi-Center Contrastive Learning (MCCL) is a prototype-based contrastive learning mechanism introduced within the Hierarchical Identity Learning (HIL) framework for unsupervised visible-infrared person re-identification (USVI-ReID). In that setting, MCCL replaces the conventional single-center representation of each pseudo-identity with multiple fine-grained memories obtained by secondary clustering inside each coarse cluster, and uses these multiple centers to construct intra-modal and cross-modal contrastive objectives. Its stated purpose is to enhance intra-modal clustering, reduce cross-modal discrepancies, and provide richer positive supervision and more informative negatives than single-prototype approaches (Shi et al., 15 Sep 2025).

1. Conceptual basis

USVI-ReID aims to learn modality-invariant image features from unlabeled cross-modal person datasets by reducing the modality gap while minimizing reliance on costly manual annotations. Existing methods are described as typically addressing this problem with cluster-based contrastive learning in which a person is represented by a single cluster center. The limitation identified by HIL is that a single-center representation emphasizes the commonality of images within a cluster while neglecting finer-grained intra-identity variation such as pose, viewpoint, occlusion, or style differences. The paper further argues that this limitation is amplified in cross-modality settings, because visible-versus-infrared discrepancies compound intra-class diversity (Shi et al., 15 Sep 2025).

MCCL addresses that limitation through a hierarchical identity model. After primary coarse-grained clustering, each coarse identity cluster is partitioned by a secondary clustering procedure into multiple fine-grained sub-clusters. These sub-clusters function as fine-grained memories within a coarse identity. MCCL then uses those multiple centers to pull each instance toward several fine-grained centers from the same coarse identity and to align instances across modalities with multi-center prototypes rather than a single coarse centroid. This suggests a shift from coarse prototype supervision toward a representation in which intra-identity structure is explicitly modeled rather than treated as noise (Shi et al., 15 Sep 2025).

2. Hierarchical representation and memory construction

The HIL formulation starts from an unlabeled bimodal dataset

$\mathcal{X} = \{\mathcal{X}^v,\mathcal{X}^r\},$

where

$$\mathcal{X}^v=\{x_k^v\}_{k=1}^{N_v}, \qquad \mathcal{X}^r=\{x_k^r\}_{k=1}^{N_r}.$$

A shared encoder $f_\theta$ maps images to $d$-dimensional features,

$$\mathbf{f}_k^v=f_\theta(x_k^v)\in\mathbb{R}^d,\qquad \mathbf{f}_k^r=f_\theta(x_k^r)\in\mathbb{R}^d.$$

Primary clustering is performed separately on visible and infrared features with DBSCAN, producing $M_v$ visible coarse clusters and $M_r$ infrared coarse clusters. For cluster $i$, the coarse prototype is the mean feature of the cluster: $$\mathbf{u}_i^{vc}=\frac{1}{|\mathcal{H}_i^{vc}|}\sum_{\mathbf{f}_n^{vc}\in\mathcal{H}_i^{vc}}\mathbf{f}_n^{vc},\qquad \mathbf{u}_i^{rc}=\frac{1}{|\mathcal{H}_i^{rc}|}\sum_{\mathbf{f}_n^{rc}\in\mathcal{H}_i^{rc}}\mathbf{f}_n^{rc}.$$ Within each coarse cluster, secondary clustering then applies $K$-means to produce $$\mathcal{X}^v=\{x_k^v\}_{k=1}^{N_v}, \qquad \mathcal{X}^r=\{x_k^r\}_{k=1}^{N_r}.$$0 fine-grained sub-clusters per identity, with fine-grained prototypes

$$\mathcal{X}^v=\{x_k^v\}_{k=1}^{N_v}, \qquad \mathcal{X}^r=\{x_k^r\}_{k=1}^{N_r}.$$1

For compactness in MCCL, the paper notes that one can view $$\mathcal{X}^v=\{x_k^v\}_{k=1}^{N_v}, \qquad \mathcal{X}^r=\{x_k^r\}_{k=1}^{N_r}.$$2 as the $$\mathcal{X}^v=\{x_k^v\}_{k=1}^{N_v}, \qquad \mathcal{X}^r=\{x_k^r\}_{k=1}^{N_r}.$$3-th fine-grained prototype in coarse cluster $$\mathcal{X}^v=\{x_k^v\}_{k=1}^{N_v}, \qquad \mathcal{X}^r=\{x_k^r\}_{k=1}^{N_r}.$$4 in the corresponding modality-specific memory bank (Shi et al., 15 Sep 2025).

Prototype maintenance follows two schedules. Coarse pseudo-labels are recomputed at the start of each epoch using DBSCAN, and secondary $$\mathcal{X}^v=\{x_k^v\}_{k=1}^{N_v}, \qquad \mathcal{X}^r=\{x_k^r\}_{k=1}^{N_r}.$$5-means is then run within each coarse cluster. Coarse prototypes are updated during training with momentum,

$$\mathcal{X}^v=\{x_k^v\}_{k=1}^{N_v}, \qquad \mathcal{X}^r=\{x_k^r\}_{k=1}^{N_r}.$$6

with default $$\mathcal{X}^v=\{x_k^v\}_{k=1}^{N_v}, \qquad \mathcal{X}^r=\{x_k^r\}_{k=1}^{N_r}.$$7. Fine-grained prototypes are initialized by $$\mathcal{X}^v=\{x_k^v\}_{k=1}^{N_v}, \qquad \mathcal{X}^r=\{x_k^r\}_{k=1}^{N_r}.$$8-means centroids each epoch and are updated during training; the implementation updates both coarse and fine prototypes after loss computation. Pseudo-labels are recomputed at the start of each epoch, and every instance is assigned a sub-cluster index $$\mathcal{X}^v=\{x_k^v\}_{k=1}^{N_v}, \qquad \mathcal{X}^r=\{x_k^r\}_{k=1}^{N_r}.$$9 corresponding to one of the $f_\theta$0 sub-centers. The fixed value of $f_\theta$1 is dataset-specific: $f_\theta$2 for SYSU-MM01 and $f_\theta$3 for RegDB (Shi et al., 15 Sep 2025).

3. Objective function and contrastive mechanism

MCCL uses cosine similarity,

$f_\theta$4

and a temperature $f_\theta$5, with default $f_\theta$6. For a visible anchor $f_\theta$7, the first step is to find the most similar visible fine-grained prototype: $f_\theta$8 If $f_\theta$9 denotes the coarse index of that best-matching sub-center, the positive set is all $d$0 fine-grained prototypes inside the same coarse cluster,

$d$1

while the negative set contains one representative fine-grained prototype from each other coarse cluster, chosen as the most similar sub-center within that cluster: $d$2 The corresponding multi-positive InfoNCE-style loss is

$d$3

$d$4

An analogous loss $d$5 is defined for infrared anchors. The paper specifies that positives are uniformly weighted and that no explicit center weighting is used (Shi et al., 15 Sep 2025).

Cross-modal MCCL follows the same construction but searches the other modality’s fine-grained memory bank. For visible-to-infrared, a visible anchor selects the most similar infrared fine-grained prototype; positives are all infrared sub-centers within that prototype’s coarse identity, and negatives are one infrared sub-center from every other coarse identity. The reverse infrared-to-visible term is defined symmetrically. The cross-modal contribution is

$d$6

and the total MCCL term is

$d$7

Within HIL, MCCL is combined with a coarse identity InfoNCE loss and a neighbor-instance contrastive loss. The coarse identity term is

$d$8

and the total training loss is

$d$9

with default $$\mathbf{f}_k^v=f_\theta(x_k^v)\in\mathbb{R}^d,\qquad \mathbf{f}_k^r=f_\theta(x_k^r)\in\mathbb{R}^d.$$0 and $$\mathbf{f}_k^v=f_\theta(x_k^v)\in\mathbb{R}^d,\qquad \mathbf{f}_k^r=f_\theta(x_k^r)\in\mathbb{R}^d.$$1. In this design, positives are all fine-grained prototypes from the anchor’s most similar coarse identity, and negatives are representative fine-grained prototypes from all other coarse identities, selected by maximal similarity to the anchor. A plausible implication is that MCCL uses coarse identity discovery to organize supervision while preserving fine-grained intra-identity structure during contrastive optimization (Shi et al., 15 Sep 2025).

4. Interaction with BRST and the HIL training pipeline

MCCL is coupled in HIL with Bidirectional Reverse Selection Transmission (BRST), a cross-modal pseudo-label association mechanism intended to establish reliable visible-to-infrared and infrared-to-visible correspondences. BRST constructs a similarity matrix

$$\mathbf{f}_k^v=f_\theta(x_k^v)\in\mathbb{R}^d,\qquad \mathbf{f}_k^r=f_\theta(x_k^r)\in\mathbb{R}^d.$$2

between visible instances and infrared coarse prototypes. For each visible instance $$\mathbf{f}_k^v=f_\theta(x_k^v)\in\mathbb{R}^d,\qquad \mathbf{f}_k^r=f_\theta(x_k^r)\in\mathbb{R}^d.$$3, it computes

$$\mathbf{f}_k^v=f_\theta(x_k^v)\in\mathbb{R}^d,\qquad \mathbf{f}_k^r=f_\theta(x_k^r)\in\mathbb{R}^d.$$4

then performs a reverse check on column $$\mathbf{f}_k^v=f_\theta(x_k^v)\in\mathbb{R}^d,\qquad \mathbf{f}_k^r=f_\theta(x_k^r)\in\mathbb{R}^d.$$5,

$$\mathbf{f}_k^v=f_\theta(x_k^v)\in\mathbb{R}^d,\qquad \mathbf{f}_k^r=f_\theta(x_k^r)\in\mathbb{R}^d.$$6

The visible instance is assigned the infrared coarse label $$\mathbf{f}_k^v=f_\theta(x_k^v)\in\mathbb{R}^d,\qquad \mathbf{f}_k^r=f_\theta(x_k^r)\in\mathbb{R}^d.$$7 if

$$\mathbf{f}_k^v=f_\theta(x_k^v)\in\mathbb{R}^d,\qquad \mathbf{f}_k^r=f_\theta(x_k^r)\in\mathbb{R}^d.$$8

The threshold is $$\mathbf{f}_k^v=f_\theta(x_k^v)\in\mathbb{R}^d,\qquad \mathbf{f}_k^r=f_\theta(x_k^r)\in\mathbb{R}^d.$$9 for SYSU-MM01 and $M_v$0 for RegDB. A symmetric procedure labels infrared instances against visible coarse prototypes, and BRST alternates direction across epochs: visible receives labels at even epochs, infrared at odd epochs. The stated interaction is that BRST filters unreliable cross-modal matches and thereby improves the quality of the cross-modal pseudo-labels used in $M_v$1 and $M_v$2, reducing modality discrepancy and stabilizing multi-center alignment (Shi et al., 15 Sep 2025).

The training pipeline uses a TransReID backbone within a dual-path transformer architecture, as in SDCL. SDCL is pretrained for 30 epochs, followed by 30 epochs of HIL training. In each epoch, the model extracts features for all visible and infrared training images, runs DBSCAN to obtain coarse pseudo-labels, computes coarse centroids, performs secondary $M_v$3-means inside each coarse identity to obtain fine-grained centroids, applies BRST label transmission, and then iterates over minibatches. Each minibatch computes $M_v$4, $M_v$5, and $M_v$6, combines them into $M_v$7, updates network parameters with SGD, and updates coarse and fine-grained centers. The optimizer is SGD with initial learning rate $M_v$8, decayed by $M_v$9 at epochs 20 and 40; the input size is $M_r$0; batches contain 8 pseudo-identities and 16 instances per pseudo-identity per modality; and the DBSCAN hyperparameters are tuned, with best SYSU-MM01 performance reported around $M_r$1 and $M_r$2 (Shi et al., 15 Sep 2025).

5. Empirical behavior, computational profile, and reproducibility

The empirical evaluation is reported on SYSU-MM01 and RegDB, using CMC, mAP, and mINP. On SYSU-MM01, training uses 22,258 visible and 11,909 infrared images of 395 identities; testing uses 96 identities with 3,803 infrared query images and a visible gallery of 301 randomly sampled images, with averages over 10 random galleries for All Search and Indoor Search. On RegDB, the dataset contains 412 identities with 10 visible and 10 infrared images per identity; 206 identities are used for training and 206 for testing, and results are averaged over 10 random splits in both visible-to-infrared and infrared-to-visible settings (Shi et al., 15 Sep 2025).

MCCL’s ablation results isolate its contribution within HIL. On SYSU-MM01 (All Search), adding MCCL to the baseline yields $M_r$3 Rank-1 and $M_r$4 mAP. On RegDB (Visible-to-Infrared), MCCL yields $M_r$5 Rank-1 and $M_r$6 mAP. With BRST already enabled, MCCL further improves SYSU-MM01 (All Search) by $M_r$7 Rank-1 and $M_r$8 mAP, and RegDB (Visible-to-Infrared) by $M_r$9 Rank-1 and $i$0 mAP. The paper summarizes this as evidence that multi-center prototypes capture intra-class nuances and provide richer positives and better negatives than single-center baselines (Shi et al., 15 Sep 2025).

Benchmark	Protocol	Performance
SYSU-MM01	All Search	Rank-1 $i$1, mAP $i$2, mINP $i$3
SYSU-MM01	Indoor Search	Rank-1 $i$4, mAP $i$5, mINP $i$6
RegDB	Visible-to-Infrared	Rank-1 $i$7, mAP $i$8, mINP $i$9
RegDB	Infrared-to-Visible	Rank-1 $$\mathbf{u}_i^{vc}=\frac{1}{|\mathcal{H}_i^{vc}|}\sum_{\mathbf{f}_n^{vc}\in\mathcal{H}_i^{vc}}\mathbf{f}_n^{vc},\qquad \mathbf{u}_i^{rc}=\frac{1}{|\mathcal{H}_i^{rc}|}\sum_{\mathbf{f}_n^{rc}\in\mathcal{H}_i^{rc}}\mathbf{f}_n^{rc}.$$0, mAP $$\mathbf{u}_i^{vc}=\frac{1}{|\mathcal{H}_i^{vc}|}\sum_{\mathbf{f}_n^{vc}\in\mathcal{H}_i^{vc}}\mathbf{f}_n^{vc},\qquad \mathbf{u}_i^{rc}=\frac{1}{|\mathcal{H}_i^{rc}|}\sum_{\mathbf{f}_n^{rc}\in\mathcal{H}_i^{rc}}\mathbf{f}_n^{rc}.$$1, mINP $$\mathbf{u}_i^{vc}=\frac{1}{|\mathcal{H}_i^{vc}|}\sum_{\mathbf{f}_n^{vc}\in\mathcal{H}_i^{vc}}\mathbf{f}_n^{vc},\qquad \mathbf{u}_i^{rc}=\frac{1}{|\mathcal{H}_i^{rc}|}\sum_{\mathbf{f}_n^{rc}\in\mathcal{H}_i^{rc}}\mathbf{f}_n^{rc}.$$2

The computational profile is explicitly characterized. Let $$\mathbf{u}_i^{vc}=\frac{1}{|\mathcal{H}_i^{vc}|}\sum_{\mathbf{f}_n^{vc}\in\mathcal{H}_i^{vc}}\mathbf{f}_n^{vc},\qquad \mathbf{u}_i^{rc}=\frac{1}{|\mathcal{H}_i^{rc}|}\sum_{\mathbf{f}_n^{rc}\in\mathcal{H}_i^{rc}}\mathbf{f}_n^{rc}.$$3 be the number of instances, $$\mathbf{u}_i^{vc}=\frac{1}{|\mathcal{H}_i^{vc}|}\sum_{\mathbf{f}_n^{vc}\in\mathcal{H}_i^{vc}}\mathbf{f}_n^{vc},\qquad \mathbf{u}_i^{rc}=\frac{1}{|\mathcal{H}_i^{rc}|}\sum_{\mathbf{f}_n^{rc}\in\mathcal{H}_i^{rc}}\mathbf{f}_n^{rc}.$$4 the number of coarse clusters, $$\mathbf{u}_i^{vc}=\frac{1}{|\mathcal{H}_i^{vc}|}\sum_{\mathbf{f}_n^{vc}\in\mathcal{H}_i^{vc}}\mathbf{f}_n^{vc},\qquad \mathbf{u}_i^{rc}=\frac{1}{|\mathcal{H}_i^{rc}|}\sum_{\mathbf{f}_n^{rc}\in\mathcal{H}_i^{rc}}\mathbf{f}_n^{rc}.$$5 the number of sub-centers per cluster, and $$\mathbf{u}_i^{vc}=\frac{1}{|\mathcal{H}_i^{vc}|}\sum_{\mathbf{f}_n^{vc}\in\mathcal{H}_i^{vc}}\mathbf{f}_n^{vc},\qquad \mathbf{u}_i^{rc}=\frac{1}{|\mathcal{H}_i^{rc}|}\sum_{\mathbf{f}_n^{rc}\in\mathcal{H}_i^{rc}}\mathbf{f}_n^{rc}.$$6 the feature dimension. BRST has space complexity $$\mathbf{u}_i^{vc}=\frac{1}{|\mathcal{H}_i^{vc}|}\sum_{\mathbf{f}_n^{vc}\in\mathcal{H}_i^{vc}}\mathbf{f}_n^{vc},\qquad \mathbf{u}_i^{rc}=\frac{1}{|\mathcal{H}_i^{rc}|}\sum_{\mathbf{f}_n^{rc}\in\mathcal{H}_i^{rc}}\mathbf{f}_n^{rc}.$$7 and time complexity $$\mathbf{u}_i^{vc}=\frac{1}{|\mathcal{H}_i^{vc}|}\sum_{\mathbf{f}_n^{vc}\in\mathcal{H}_i^{vc}}\mathbf{f}_n^{vc},\qquad \mathbf{u}_i^{rc}=\frac{1}{|\mathcal{H}_i^{rc}|}\sum_{\mathbf{f}_n^{rc}\in\mathcal{H}_i^{rc}}\mathbf{f}_n^{rc}.$$8. MCCL has space complexity $$\mathbf{u}_i^{vc}=\frac{1}{|\mathcal{H}_i^{vc}|}\sum_{\mathbf{f}_n^{vc}\in\mathcal{H}_i^{vc}}\mathbf{f}_n^{vc},\qquad \mathbf{u}_i^{rc}=\frac{1}{|\mathcal{H}_i^{rc}|}\sum_{\mathbf{f}_n^{rc}\in\mathcal{H}_i^{rc}}\mathbf{f}_n^{rc}.$$9 and time complexity $K$0. The full model has 98.61M parameters and 32.78 GFLOPs per image at inference. The official implementation is released at the HIL repository. The practical guidance provided in the paper emphasizes that larger $K$1 increases granularity and cost, too small $K$2 may underfit intra-class variation, $K$3 works well but lower values may cause instability, and $K$4 provides the best validated balance among losses (Shi et al., 15 Sep 2025).

6. Terminological scope, limitations, and future directions

The acronym MCCL is not unique across recent arXiv literature, and this has produced a recurrent source of confusion. In "SynSeg," MCCL denotes Multi-Category Contrastive Learning, a weakly supervised open-vocabulary semantic segmentation objective that combines intra-category alignment and separation with inter-category background alignment and foreground separation; the paper explicitly states that it does not use learnable centers (Zhang et al., 8 Aug 2025). In "MICACL," MCCL denotes Multiscale Category-aware Contrastive Learning, an in-batch supervised contrastive objective for long-tailed dynamic facial expression recognition that uses multiscale feature heads, category-aware weighting, and a class-frequency-dependent temperature, and explicitly states that it does not introduce class centers, proxies, or multi-center prototypes (Cui et al., 4 Sep 2025). In "Memory-aided Contrastive Consensus Learning," MCCL denotes a co-salient object detection framework with a memory-based contrastive module and triplet-style loss; that work states that it is not a multi-center or multi-prototype contrastive scheme in the classical sense (Zheng et al., 2023). A separate but related line appears in "A Contrastive Pretrain Model with Prompt Tuning for Multi-center Medication Recommendation," where contrastive learning operates in a multi-hospital setting and “multi-center” refers to hospitals and cross-center distributional heterogeneity rather than multi-prototype representation learning (Liu et al., 2024). This suggests that, in current usage, the acronym MCCL is context-dependent and should not be interpreted as a stable reference to a single methodology.

For the specific multi-center prototype formulation in HIL, the stated limitations are sensitivity to clustering errors, modality imbalance or bias, and the use of a fixed $K$5 per identity. If coarse or secondary clustering is poor, pseudo-label quality degrades and therefore so do MCCL positives and negatives. If one modality clusters more weakly than the other, cross-modal alignment may be affected. A fixed $K$6 may also be suboptimal when cluster size and density vary across identities. The future directions identified by the authors are adaptive thresholding and dynamic refinement in BRST, adaptive $K$7 based on cluster statistics, robust clustering under modality imbalance, and weighting schemes over sub-centers in MCCL. In that sense, the current formulation establishes a hierarchical prototype framework, while leaving open the question of how prototype multiplicity should itself be adapted to data geometry (Shi et al., 15 Sep 2025).