PCSR: Pseudo-label Consistency-Guided Refinement

• Pseudo-label Consistency-Guided Sample Refinement is a training strategy family that refines noisy pseudo labels using explicit consistency tests like temporal agreement and uncertainty weighting.
• It uses dynamic mechanisms—such as cross-generation clustering, class-discriminative filtering, and teacher consensus—to selectively propagate reliable pseudo supervision.
• Applications include unsupervised re-identification, weakly supervised localization, breast ultrasound segmentation, and cross-modal retrieval with demonstrated performance improvements.

Pseudo-label Consistency-Guided Sample Refinement (PCSR) denotes a family of training strategies in which pseudo supervision is not used as a static target, but is filtered, propagated, fused, or reweighted according to a notion of consistency. In the supplied literature, the term is explicit in noisy correspondence learning for cross-modal retrieval, while closely related mechanisms appear in unsupervised object re-identification as RLCC, in weakly supervised object localization as LCHP’s high-quality pseudo-label consistency learning, and in breast ultrasound segmentation as a PCSR-like refinement pipeline based on teacher fusion and uncertainty control (Liu et al., 19 Sep 2025, Zhang et al., 2021, Sun et al., 2022, Li et al., 6 Mar 2026). Across these settings, the common objective is to reduce the damage caused by noisy or unstable pseudo labels by exploiting temporal agreement, augmentation consistency, classification discrimination, or uncertainty-weighted teacher agreement.

1. Conceptual basis and problem setting

PCSR addresses a recurrent failure mode of pseudo-supervised learning: pseudo labels are useful only insofar as their noise is controlled. The cited works instantiate this problem in different domains. In clustering-based unsupervised re-identification, pseudo identities are regenerated after each epoch and are therefore noisy and unstable across generations (Zhang et al., 2021). In weakly supervised object localization, pseudo bounding boxes derived from class labels may cover only discriminative object parts or include excessive background, so refinement depends on selecting high-quality pseudo boxes (Sun et al., 2022). In semi-supervised breast ultrasound segmentation, pseudo masks generated under extremely limited annotation can be unstable, especially near weak and speckle-corrupted lesion boundaries (Li et al., 6 Mar 2026). In cross-modal retrieval, the problem is not incorrect class labels but noisy correspondence: an image-caption pair may be semantically mismatched while still being treated as a positive pair during training (Liu et al., 19 Sep 2025).

A central misconception is that pseudo-label refinement is equivalent to indiscriminate smoothing. The supplied works instead treat consistency as a selective criterion. In RLCC for unsupervised re-identification, previous pseudo labels are retained only after cross-generation clustering consensus maps old classes to new ones (Zhang et al., 2021). In LCHP for weakly supervised localization, only samples whose cropped box is sufficiently class-discriminative contribute to the refinement loss (Sun et al., 2022). In BUS segmentation, uncertainty–entropy weighted fusion suppresses high-entropy teacher outputs at each spatial location rather than averaging teachers uniformly (Li et al., 6 Mar 2026). In noisy correspondence learning, PCSR does not merely divide data into clean and noisy pairs; it further separates noisy pairs into ambiguous and refinable subsets via a Pseudo-label Consistency Score (PCS), and each subset is optimized differently (Liu et al., 19 Sep 2025).

This suggests that PCSR is best understood not as a single algorithm, but as a methodological pattern: pseudo labels are promoted to reliable supervision only after an explicit consistency test.

2. Generic refinement logic

Despite domain differences, the supplied methods share a structurally similar pipeline. First, pseudo supervision is produced by a base mechanism: clustering in unsupervised re-identification, mask-based box generation in weakly supervised localization, VLM prompting plus Grounding DINO and SAM in BUS segmentation, or confidence-based pair separation in cross-modal retrieval (Zhang et al., 2021, Sun et al., 2022, Li et al., 6 Mar 2026, Liu et al., 19 Sep 2025). Second, a consistency or reliability signal is computed. Third, that signal is used to refine supervision before subsequent optimization.

The main refinement signals differ:

Setting	Pseudo-label source	Consistency or reliability signal
UReID / RLCC	DBSCAN pseudo identities	Cross-generation cluster IoU consensus
WSOL / LCHP	Mask-based pseudo boxes	Classification discrimination of cropped box
BUS segmentation	VLM-generated masks; dual teachers	Shannon-entropy-based uncertainty weighting
Cross-modal retrieval / PCSR	GMM clean/noisy split; pseudo-classifier	PCS from repeated pseudo-label predictions

In RLCC, the consistency signal is a consensus matrix between pseudo classes from consecutive generations. If $\mathcal{I}^{(t-1)}(i)$ and $\mathcal{I}^{(t)}(j)$ denote the samples assigned to pseudo classes $i$ and $j$ at generations $t-1$ and $t$, the clustering consensus is defined as

$$\mathbf{C}^{(t-1,t)}(i,j)=\frac{\left|\mathcal{I}^{(t-1)}(i)\cap \mathcal{I}^{(t)}(j)\right|}{\left|\mathcal{I}^{(t-1)}(i)\cup \mathcal{I}^{(t)}(j)\right|}\in[0,1].$$

The row-normalized matrix acts like a transition matrix from old pseudo classes to current ones, enabling temporal propagation even though the label spaces differ across epochs (Zhang et al., 2021).

In LCHP, reliability is defined by whether the predicted crop remains class-discriminative. If $I_b$ is the cropped image inside the predicted box and $y^{cls}_b$ is its classification distribution, a pseudo box is retained only when $\max(y^{cls}_{b})>\tau$ (Sun et al., 2022). In the BUS segmentation framework, teacher predictions are converted into pixel-wise entropy maps

$\mathcal{I}^{(t)}(j)$0

smoothed with average pooling and upsampling, and then inverted into confidence weights for fusion (Li et al., 6 Mar 2026). In the cross-modal PCSR framework, prediction stability across multiple epochs is summarized by

$\mathcal{I}^{(t)}(j)$1

where $\mathcal{I}^{(t)}(j)$2 and $\mathcal{I}^{(t)}(j)$3 are the largest and second-largest counts in the pseudo-class histogram for sample $\mathcal{I}^{(t)}(j)$4 (Liu et al., 19 Sep 2025).

A plausible implication is that PCSR methods differ less in their downstream loss functions than in how they decide which pseudo supervision should be trusted.

3. RLCC in unsupervised object re-identification

The 2021 re-identification method formalizes the PCSR idea under the name RLCC, “Refining pseudo Labels with Clustering Consensus over generations” (Zhang et al., 2021). It is introduced for clustering-based unsupervised object re-identification, where methods such as BUC, HCT, and SpCL repeatedly extract features, cluster them into pseudo identities, train on those pseudo labels, and repeat. The stated bottleneck is pseudo-label noise, arising because class identities are discovered from an evolving representation and therefore change from epoch to epoch (Zhang et al., 2021).

The method’s central novelty is cross-generation label propagation. After re-clustering at generation $\mathcal{I}^{(t)}(j)$5 with DBSCAN, the consensus matrix between generations $\mathcal{I}^{(t)}(j)$6 and $\mathcal{I}^{(t)}(j)$7 is computed via the IoU overlap above and row-normalized:

$\mathcal{I}^{(t)}(j)$8

This matrix provides the “pseudo-label consistency guidance.” Because classes at $\mathcal{I}^{(t)}(j)$9 and $i$0 are different, ordinary temporal ensembling is not directly applicable. RLCC therefore propagates supervision through the consensus matrix. For hard labels,

$i$1

and for soft pseudo-label confidence propagation,

$i$2

The paper identifies soft propagation as the preferred variant because it is more informative than one-hot labels and empirically performs better (Zhang et al., 2021).

The propagated target is then combined with the current generation’s pseudo label through momentum averaging:

$i$3

Training uses the refined soft pseudo labels in a cross-entropy objective:

$i$4

For SpCL, the framework adapts to a non-parametric memory bank with dynamic class prototypes, and the paper gives the prototype-based soft propagation form

$i$5

where $i$6 are previous normalized prototypes and $i$7 is a sharpening temperature (Zhang et al., 2021).

The reported results show consistent improvement over SpCL: RLCC reaches 77.7% mAP on Market-1501, 69.2% mAP on DukeMTMC-reID, and 27.9% mAP on MSMT17, corresponding to gains of 4.6%, 3.9%, and 8.8% mAP respectively; on VeRi-776 the gain is 2.7% mAP (Zhang et al., 2021). The ablations further report that soft propagation outperforms hard propagation across tested momentum values, the best momentum is approximately $i$8, the best temperature is $i$9, and useful robustness is observed for $j$0 and $j$1 (Zhang et al., 2021). These findings are consistent with the paper’s interpretation that moderate temporal smoothing can stabilize optimization without erasing newly discovered identity structure.

4. High-quality pseudo-label selection in weakly supervised object localization

LCHP, “Learning Consistency from High-quality Pseudo-labels,” applies a related refinement principle to weakly supervised object localization, where only image-level class labels are available and the standard assumption is that there is one instance per image (Sun et al., 2022). The method is explicitly two-stage. The initialization stage, LCHP-I, generates pseudo bounding boxes and trains a localization network. The refinement stage, LCHP-R, estimates which pseudo boxes are reliable and applies consistency regularization only to those high-quality pseudo labels (Sun et al., 2022).

The pseudo-label generator begins from a raw image $j$2. CNN feature maps $j$3 are averaged channel-wise to obtain an attention map $j$4, thresholded by $j$5, and converted into a smallest enclosing rectangle $j$6 over the foreground pixels. Because this region often captures only the most salient object part, the method masks out $j$7 in the original image to produce $j$8, repeats localization once on $j$9 to obtain $t-1$0, and merges $t-1$1 and $t-1$2 into the pseudo box $t-1$3 (Sun et al., 2022). The stage-one losses are

$t-1$4

$t-1$5

and

$t-1$6

The refinement stage defines pseudo-label quality through classification discrimination. For a predicted box, the cropped image $t-1$7 is classified, and the box is retained only if the maximum class probability exceeds the threshold $t-1$8, represented by the indicator $t-1$9 (Sun et al., 2022). Consistency learning is then imposed between a raw image and its strong augmentation. If $t$0 and $t$1 is a strong augmentation applied jointly to image and box, then

$t$2

The basic refinement objective is

$t$3

and the high-quality filtered version is

$t$4

Only pseudo boxes judged sufficiently reliable are allowed to regularize the localization network (Sun et al., 2022).

The reported experiments show consistent gains from this refinement. On CUB-200-2011, the baseline pseudo-label method yields 61.03% Top-1 and 72.44% GT-Known, LCHP-I reaches 73.61% Top-1 and 87.12% GT-Known, and LCHP-R reaches 80.39% Top-1 and 94.36% GT-Known (Sun et al., 2022). On ImageNet-1k, results progress from 47.21% and 57.57% to 52.59% and 64.11%, then to 54.12% and 66.08%; on Tiny-ImageNet, from 41.24% and 50.33% to 49.06% and 59.80%, then to 50.95% and 61.87% (Sun et al., 2022). The ablation on $t$5 is especially informative: keeping all boxes by setting $t$6 can cause refinement to fail, while increasing $t$7 improves GT-Known performance, with the best CUB result occurring at $t$8 (Sun et al., 2022). This directly supports the claim that pseudo-label consistency is beneficial only when preceded by quality control.

5. PCSR-like refinement in semi-supervised breast ultrasound segmentation

The 2026 breast ultrasound segmentation framework does not name its refinement module PCSR, but the supplied extraction identifies a PCSR-like system composed of Appearance-Prompted Training-Free Pseudo-Label Generation (APPG), static-teacher warm-up, uncertainty–entropy weighted fusion (UEWF), adaptive uncertainty-guided reverse contrastive learning (AURCL), and mask filtering (Li et al., 6 Mar 2026). The setting is semi-supervised BUS segmentation under extremely limited labels, where pseudo masks can be unstable and diagnosis-oriented prompts transfer poorly from natural-image VLMs (Li et al., 6 Mar 2026).

APPG uses appearance-based prompts derived from general lesion descriptors:

$t$9

with prompt examples including “dark oval,” “dark round,” and “dark lobulated” (Li et al., 6 Mar 2026). For an unlabeled image $$\mathbf{C}^{(t-1,t)}(i,j)=\frac{\left|\mathcal{I}^{(t-1)}(i)\cap \mathcal{I}^{(t)}(j)\right|}{\left|\mathcal{I}^{(t-1)}(i)\cup \mathcal{I}^{(t)}(j)\right|}\in[0,1].$$0, the initial pseudo label is

$$\mathbf{C}^{(t-1,t)}(i,j)=\frac{\left|\mathcal{I}^{(t-1)}(i)\cap \mathcal{I}^{(t)}(j)\right|}{\left|\mathcal{I}^{(t-1)}(i)\cup \mathcal{I}^{(t)}(j)\right|}\in[0,1].$$1

The implementation is specifically two-step: Grounding DINO first localizes lesion-like regions,

$$\mathbf{C}^{(t-1,t)}(i,j)=\frac{\left|\mathcal{I}^{(t-1)}(i)\cap \mathcal{I}^{(t)}(j)\right|}{\left|\mathcal{I}^{(t-1)}(i)\cup \mathcal{I}^{(t)}(j)\right|}\in[0,1].$$2

and SAM converts the predicted box into a mask,

$$\mathbf{C}^{(t-1,t)}(i,j)=\frac{\left|\mathcal{I}^{(t-1)}(i)\cap \mathcal{I}^{(t)}(j)\right|}{\left|\mathcal{I}^{(t-1)}(i)\cup \mathcal{I}^{(t)}(j)\right|}\in[0,1].$$3

These masks are filtered so that only those with $$\mathbf{C}^{(t-1,t)}(i,j)=\frac{\left|\mathcal{I}^{(t-1)}(i)\cap \mathcal{I}^{(t)}(j)\right|}{\left|\mathcal{I}^{(t-1)}(i)\cup \mathcal{I}^{(t)}(j)\right|}\in[0,1].$$4 are retained, with $$\mathbf{C}^{(t-1,t)}(i,j)=\frac{\left|\mathcal{I}^{(t-1)}(i)\cap \mathcal{I}^{(t)}(j)\right|}{\left|\mathcal{I}^{(t-1)}(i)\cup \mathcal{I}^{(t)}(j)\right|}\in[0,1].$$5 of the image area (Li et al., 6 Mar 2026).

Refinement then proceeds through two teachers. A static teacher $$\mathbf{C}^{(t-1,t)}(i,j)=\frac{\left|\mathcal{I}^{(t-1)}(i)\cap \mathcal{I}^{(t)}(j)\right|}{\left|\mathcal{I}^{(t-1)}(i)\cup \mathcal{I}^{(t)}(j)\right|}\in[0,1].$$6 is warmed up on the valid VLM pseudo labels and frozen, so that it captures “coarse structural priors of lesion regions.” A dynamic teacher $$\mathbf{C}^{(t-1,t)}(i,j)=\frac{\left|\mathcal{I}^{(t-1)}(i)\cap \mathcal{I}^{(t)}(j)\right|}{\left|\mathcal{I}^{(t-1)}(i)\cup \mathcal{I}^{(t)}(j)\right|}\in[0,1].$$7 is updated by exponential moving average of the student parameters, $$\mathbf{C}^{(t-1,t)}(i,j)=\frac{\left|\mathcal{I}^{(t-1)}(i)\cap \mathcal{I}^{(t)}(j)\right|}{\left|\mathcal{I}^{(t-1)}(i)\cup \mathcal{I}^{(t)}(j)\right|}\in[0,1].$$8 (Li et al., 6 Mar 2026). For each unlabeled image, the two teachers produce soft pseudo labels $$\mathbf{C}^{(t-1,t)}(i,j)=\frac{\left|\mathcal{I}^{(t-1)}(i)\cap \mathcal{I}^{(t)}(j)\right|}{\left|\mathcal{I}^{(t-1)}(i)\cup \mathcal{I}^{(t)}(j)\right|}\in[0,1].$$9 and $I_b$0. Entropy maps are smoothed by

$I_b$1

with $I_b$2, converted into confidence weights

$I_b$3

and fused into

$I_b$4

This refinement is explicitly uncertainty-aware: low entropy yields high confidence and therefore higher influence in the fused pseudo label (Li et al., 6 Mar 2026).

AURCL complements fusion by targeting uncertain pixels rather than ignoring them. For student prediction $I_b$5, confidence is defined as $I_b$6, and a dynamic threshold

$I_b$7

with $I_b$8, identifies low-confidence pixels via

$I_b$9

On those pixels, probabilities are reversed:

$y^{cls}_b$0

and patch-level features are contrasted with the AURCL loss in Eq. (15) of the supplied data (Li et al., 6 Mar 2026).

The total training objective is

$y^{cls}_b$1

with $y^{cls}_b$2 and $y^{cls}_b$3, where $y^{cls}_b$4 is BCE + Dice on labeled data, $y^{cls}_b$5 applies the same formulation using fused pseudo labels, and $y^{cls}_b$6 (Li et al., 6 Mar 2026). On BUSI with 2.5% labels, ablation results progress from Dice 50.00 and IoU 39.71 for supervised only, to 71.17 and 61.63 with APPG, to 71.64 and 62.49 with AURCL, to 72.16 and 62.56 with UEWF, and to 72.72 and 63.11 for the full model; on UBB with 2.5% labels, the method achieves Dice 75.75 and IoU 65.86 (Li et al., 6 Mar 2026). The prompt comparison further reports that medical terms such as “tumor,” “lesion,” “breast cancer,” and “high density” perform worse than appearance-based prompts (Li et al., 6 Mar 2026). Here, pseudo-label consistency is mediated by teacher agreement and uncertainty weighting rather than by class overlap or fixed-threshold sample acceptance.

6. Explicit PCSR in noisy correspondence learning and broader significance

The 2025 cross-modal retrieval framework is the source in which PCSR is explicitly named “Pseudo-label Consistency-Guided Sample Refinement” (Liu et al., 19 Sep 2025). The task is image-text retrieval under noisy correspondence, where web-scale data may contain image-caption pairs that do not describe the same semantic content. The paper argues that earlier approaches often rely on a coarse clean/noisy split and apply uniform strategies to all noisy samples, even though noisy pairs are heterogeneous (Liu et al., 19 Sep 2025).

PCSR begins with a GMM-based confidence division, following the noisy correspondence framework used in NCR. This yields clean pairs $y^{cls}_b$7 and noisy pairs $y^{cls}_b$8 (Liu et al., 19 Sep 2025). A pseudo-classifier $y^{cls}_b$9 is then trained only on clean pairs. For a clean batch, the classifier outputs $\max(y^{cls}_{b})>\tau$0 for images and $\max(y^{cls}_{b})>\tau$1 for texts, and is optimized with cross-entropy plus an entropy regularizer:

$\max(y^{cls}_{b})>\tau$2

$\max(y^{cls}_{b})>\tau$3

with $\max(y^{cls}_{b})>\tau$4 in the implementation (Liu et al., 19 Sep 2025).

The key refinement signal is PCS:

$\max(y^{cls}_{b})>\tau$5

computed from repeated pseudo-label predictions over training. Noisy pairs with $\max(y^{cls}_{b})>\tau$6 are assigned to the refinable subset

$\max(y^{cls}_{b})>\tau$7

while those with $\max(y^{cls}_{b})>\tau$8 become ambiguous

$\max(y^{cls}_{b})>\tau$9

The threshold is updated dynamically via a target utilization rate:

$\mathcal{I}^{(t)}(j)$00

$\mathcal{I}^{(t)}(j)$01

$\mathcal{I}^{(t)}(j)$02

with implementation values $\mathcal{I}^{(t)}(j)$03, $\mathcal{I}^{(t)}(j)$04, $\mathcal{I}^{(t)}(j)$05, and $\mathcal{I}^{(t)}(j)$06 (Liu et al., 19 Sep 2025).

Optimization is subset-specific. Clean pairs use a triplet-based objective together with the pseudo-classification losses:

$\mathcal{I}^{(t)}(j)$07

The triplet loss is

$\mathcal{I}^{(t)}(j)$08

with adaptive margin

$\mathcal{I}^{(t)}(j)$09

and implementation values $\mathcal{I}^{(t)}(j)$10 and $\mathcal{I}^{(t)}(j)$11 (Liu et al., 19 Sep 2025). Refinable pairs use caption replacement: for each refinable image $\mathcal{I}^{(t)}(j)$12, the most semantically similar clean caption is found by

$\mathcal{I}^{(t)}(j)$13

and the pair $\mathcal{I}^{(t)}(j)$14 is replaced with $\mathcal{I}^{(t)}(j)$15 and optimized through $\mathcal{I}^{(t)}(j)$16 (Liu et al., 19 Sep 2025). Ambiguous pairs are included later with a robust Generalized Cross-Entropy term,

$\mathcal{I}^{(t)}(j)$17

with $\mathcal{I}^{(t)}(j)$18, leading to

$\mathcal{I}^{(t)}(j)$19

Warm-up uses 5 epochs on Flickr30K and MSCOCO and 10 epochs on CC152K (Liu et al., 19 Sep 2025).

The reported experiments show that on CC152K, PCSR attains Rsum 375.3 compared with 374.2 for L2RM, with Image→Text R@1 of 43.7 and Text→Image R@1 of 43.1 (Liu et al., 19 Sep 2025). Under 80% synthetic noise on Flickr30K, PCSR reaches 63.1 R@1 for image→text and 44.6 for text→image, compared with 58.7 and 40.8 for CREAM (Liu et al., 19 Sep 2025). The paper further reports that $\mathcal{I}^{(t)}(j)$20 performs best in the pseudo-classifier ablation on Flickr30K with 60% noise, achieving Rsum 480.5, while using all three subsets—clean, refinable, and ambiguous—outperforms coarser alternatives (Liu et al., 19 Sep 2025). This is the most explicit statement of PCSR’s broader significance: noisy samples are not a uniform residual category, and pseudo-label consistency can be used to identify which noisy instances are recoverable rather than merely suppressible.

Taken together, these works support a unified interpretation. PCSR is a refinement doctrine in which pseudo supervision is accepted, propagated, or reused only after a measurable agreement criterion has been applied. The exact criterion may be cluster IoU across generations, class-discriminative confidence of cropped regions, uncertainty-weighted agreement between teachers, or prediction stability across training epochs. What remains constant is the assumption that pseudo labels become more useful when their internal consistency is modeled explicitly rather than presumed.