Pseudo Label Assigner (PLA)

• Pseudo Label Assigner (PLA) is a task-dependent mechanism that transforms unlabeled or weakly supervised data into trainable pseudo-labels by leveraging uncertainty and posterior criteria.
• It integrates methods such as uncertainty-aware masking, optimal transport-based allocation, and triage strategies to mitigate overconfidence and calibration errors.
• PLA plays a critical role in semi-supervised classification, object detection, and partial-label learning by converting model predictions into structured supervision.

Pseudo Label Assigner (PLA) in semi-supervised learning is the module that decides which unlabeled instances should be pseudo-labeled and added to the training set, and with which labels. In adjacent literatures, the same name or functional role also covers proposal-level assignment in semi-supervised object detection, label pairing across modalities, soft pseudo-label optimization in partial-label learning, and progressive pseudo-box generation from weak supervision. This suggests that PLA is best understood not as a single algorithm, but as a task-dependent assignment mechanism that converts incomplete, indirect, or model-generated supervision into trainable targets while attempting to limit confirmation bias, calibration error, localization noise, structural missingness, or modality misalignment (Rodemann et al., 2023, Rizve et al., 2021, Li et al., 2022, Jin et al., 14 Aug 2025, Zhang et al., 30 Sep 2025, Abdelfattah et al., 2022).

1. Conceptual scope and task-dependent outputs

At its most basic, a PLA takes labeled data $D$, unlabeled instances $U$, and model predictions, and chooses a subset of $U$ together with pseudo-labels $\hat y$ to augment training. In this form, the objective is to improve generalization without propagating early overfitting through overconfident but erroneous predictions. That formulation is explicit in Bayesian Pseudo Label Selection (BPLS), where a PLA selects candidate pseudo-samples $(x_*, \hat y_*)$ by maximizing a posterior-predictive utility (Rodemann et al., 2023).

Outside standard semi-supervised classification, the same role is instantiated more broadly. In uncertainty-aware pseudo-label selection (UPS), the PLA outputs a class-wise selection mask $g_c^{(i)}$ together with positive and negative pseudo-labels. In PseCo and Efficient Teacher, the PLA is proposal-level: it chooses which proposals or dense locations should be treated as positives, negatives, or uncertain targets for object detection. In DOD-SA, the PLA aligns labeled boxes from one modality with pseudo boxes from another modality and outputs paired supervision $y^d = \{c, b_{ir}, \tilde b_{rgb}\}$. In Point2RBox-v3, Progressive Label Assignment generates dynamic pseudo oriented boxes from point annotations. In partial-label regimes, SARI, G2NetPL, reduction-based pseudo-label generation, and active refinement use PLAs that output hard or soft class distributions over ambiguous labels rather than conventional single-label assignments (Rizve et al., 2021, Jin et al., 14 Aug 2025, Zhang et al., 30 Sep 2025, Saravanan et al., 2024, Abdelfattah et al., 2022, Qiao et al., 2024, Hsieh et al., 2021).

Setting	PLA output	Representative formulations
Semi-supervised classification	Selected pseudo-samples or class-wise masks	BPLS, UPS, CSA
Semi-supervised object detection	Positive/negative proposal assignment, pseudo boxes, objectness targets	PseCo, Efficient Teacher
Cross-modality or weakly supervised detection	Paired boxes or dynamic pseudo oriented boxes	DOD-SA, Point2RBox-v3
Partial-label and multi-label learning	Soft or hard pseudo-label distributions over candidate entries	SARI, G2NetPL, Rplg, active refinement

A common misconception is that a PLA is merely a confidence threshold. The surveyed formulations contradict that view. Some PLAs solve a Bayes decision problem, some use uncertainty-aware masking, some impose class-balance or optimal-transport constraints, and some perform structured pairing or bilevel meta-optimization. The unifying feature is assignment, not any single scoring heuristic.

2. Assignment criteria: posterior utility, uncertainty, transport, and triage

A large part of PLA research concerns the criterion used to accept or reject a pseudo-label. In BPLS, the central score is the posterior predictive of a pseudo-sample,

$$p(y_* \mid x_*, D) = \int p(y_* \mid x_*, \theta)\, p(\theta \mid D)\, d\theta,$$

and, in joint form, the pseudo posterior predictive

$p(D \cup (x_i,\hat y_i)\mid D).$

The paper proves that with the posterior prior $\pi(\theta)=p(\theta\mid D)$, maximizing this quantity is Bayes-optimal in expectation, and it derives practical Laplace-based scores

$U$0

and

$U$1

The curvature term $U$2 explicitly penalizes sharp maxima and is intended to temper early overconfidence (Rodemann et al., 2023).

UPS uses a different logic. Rather than integrating over parameter uncertainty, it filters pseudo-labels jointly by confidence and predictive uncertainty. With class probability $U$3, uncertainty $U$4, positive threshold $U$5, negative threshold $U$6, and uncertainty thresholds $U$7, the selection mask is

$U$8

This construction supports both positive and negative pseudo-labels, and the paper attributes its gains to the empirical observation that subsets filtered by low uncertainty have lower Expected Calibration Error (Rizve et al., 2021).

Confident Sinkhorn Allocation (CSA) replaces greedy thresholding by confidence-filtered optimal transport. It first applies a Welch’s T-test to ensemble predictions, accepting a sample only if the top class is statistically separated from the runner-up:

$U$9

with acceptance when $U$0. On the accepted set, it builds a cost matrix

$U$1

and then solves an entropic optimal-transport problem by Sinkhorn scaling under class-quota marginals. The PLA therefore becomes a global allocation mechanism rather than a purely local threshold rule (Nguyen et al., 2022).

Efficient Teacher introduces a triage-based PLA for dense one-stage anchor-based detection. After teacher NMS, pseudo labels are partitioned by two thresholds into reliable $U$2, uncertain $U$3, and background $U$4. Reliable pseudo labels receive standard unsupervised classification, regression, and objectness losses. Uncertain pseudo labels do not induce hard class labels; instead they provide soft objectness targets, and regression is activated only when the teacher objectness exceeds $U$5. This is a distinct assignment philosophy: uncertain predictions are neither discarded nor hardened into positives or negatives (Xu et al., 2023).

Taken together, these criteria illustrate four major design lines: posterior-expected utility, uncertainty-aware masking, global quota-constrained allocation, and multi-bin triage. None reduces to raw maximum probability alone.

3. Structured PLA in object detection and localization

In object detection, PLA is rarely a sample-level selector; it is usually a structured assigner operating on proposals, anchors, or paired boxes. PseCo’s Prediction-guided Label Assignment (PLA) is representative. For each pseudo box $U$6, it collects a candidate bag of proposals with $U$7, extracts teacher classification confidence $U$8 and proposal-wise localization quality $U$9, defines

$\hat y$0

estimates a dynamic $\hat y$1, and chooses the top-$\hat y$2 proposals as positives. The aim is to remain robust when pseudo boxes are coarse, since classification confidence alone does not guarantee localization precision (Li et al., 2022).

Efficient Teacher addresses a related but denser setting. Its PLA inherits the Dense Detector’s anchor/grid assignment but modifies how pseudo labels affect each loss branch. The objectness branch is central: reliable boxes receive hard objectness targets, uncertain boxes receive soft objectness supervision, and very high-objectness uncertain boxes can still trigger regression. This selective use of objectness is designed for one-stage anchor-based detectors, where dense predictions can otherwise amplify low-quality pseudo-label noise (Xu et al., 2023).

DOD-SA introduces a cross-modality PLA because infrared and visible boxes are rarely pixel-aligned. Its PLA has three components. The Pseudo Label Filter (PLF) retains pseudo labels whose score exceeds a batch-adaptive threshold $\hat y$3. The Shape-aware Dual-modality Label Matcher (SDLM) then matches each labeled infrared oriented box to a filtered visible pseudo box within an orientation-aligned search region and selects the best oriented-IoU match under a one-to-one constraint. Dynamic Label Correction (DLC) maintains a dynamic label pairs bag across epochs so unmatched infrared boxes are not lost and can be updated when better matches appear later. No explicit homography or alignment head is introduced; pairing is local, shape-aware, and progressively corrected (Jin et al., 14 Aug 2025).

Point2RBox-v3 uses the term PLA differently but still in the assignment sense. Progressive Label Assignment first generates pseudo oriented boxes from Watershed masks and minimum-area rectangles in early epochs, then switches at epoch $\hat y$4 to score-selected network predictions gathered across FPN levels:

$\hat y$5

These dynamic pseudo RBoxes restore multi-level label assignment under point supervision by providing coarse-to-refined size information for standard FCOS-style assignment (Zhang et al., 30 Sep 2025).

A plausible implication is that in detection, the main difficulty is not pseudo-classification alone, but structural compatibility between pseudo labels and the detector’s own assignment machinery: proposal sets, anchor grids, oriented geometry, or cross-modality correspondences.

4. PLA under partial-label, multi-label, and ambiguous supervision

When supervision is ambiguous rather than absent, PLAs must disambiguate candidate labels rather than merely accept or reject unlabeled instances. SARI does this with a weighted $\hat y$6-nearest-neighbor assigner in the learned feature space. For each class $\hat y$7,

$\hat y$8

and the initial pseudo-label is

$\hat y$9

A second score,

$(x_*, \hat y_*)$0

is used for class-balanced reliable-set selection via per-class quotas. These pseudo-labels are then used with label smoothing, mixup, and consistency regularization, and the partial-label sets are iteratively augmented by confident model predictions (Saravanan et al., 2024).

G2NetPL formulates pseudo labels as continuous decision variables in a two-player non-zero-sum non-cooperative game. For each image and class, a soft pseudo-label $(x_*, \hat y_*)$1 is optimized against the current network outputs via an augmented cross-entropy

$(x_*, \hat y_*)$2

while a confidence-aware scheduler

$(x_*, \hat y_*)$3

weights the network’s loss on unobserved labels. This PLA does not use hard thresholds; it performs easy-to-hard learning over soft labels that are pushed toward $(x_*, \hat y_*)$4 or $(x_*, \hat y_*)$5 while remaining coupled to predictions (Abdelfattah et al., 2022).

Reduction-based pseudo-label generation for instance-dependent PLL attacks a different failure mode: the predictive model itself may overfit to incorrect candidate labels. The method trains a multi-branch auxiliary model, where branch $(x_*, \hat y_*)$6 excludes label $(x_*, \hat y_*)$7, aggregates branch outputs via a meta-learner $(x_*, \hat y_*)$8, and forms

$(x_*, \hat y_*)$9

The paper proves that, under its assumptions, these reduction-based pseudo-labels are more consistent with the Bayes optimal classifier than pseudo-labels generated directly from the main predictive model (Qiao et al., 2024).

Active refinement for multi-label learning treats PLA as a bilevel meta-optimization problem. For each unknown fine-label entry $g_c^{(i)}$0, it computes the sign of

$g_c^{(i)}$1

after a one-step look-ahead update, and sets

$g_c^{(i)}$2

Here pseudo-labels are chosen explicitly to improve validation performance on a warm-up set with fully observed fine labels (Hsieh et al., 2021).

These formulations show that, in ambiguous-supervision settings, PLA often becomes a disambiguation engine over candidate labels, not a selector over unlabeled samples.

5. Integration into training loops and system architectures

PLAs are typically embedded in iterative pipelines rather than used once. In BPLS, the standard loop is: fit the initial model on $g_c^{(i)}$3, evaluate a score such as the pseudo posterior predictive or its Laplace proxy for each $g_c^{(i)}$4, select a batch by top-$g_c^{(i)}$5 or thresholding, augment $g_c^{(i)}$6, refit, and iterate. The paper notes that one-by-one selection without early stopping requires $g_c^{(i)}$7 pseudo-posterior evaluations, so approximation and local updates are operationally important (Rodemann et al., 2023).

UPS follows a similar outer loop but emphasizes recalibration and reinitialization. After supervised training, the model computes probabilities, temperature-scaled outputs, and MC-Dropout uncertainties; positives and negatives are selected by threshold pairs $g_c^{(i)}$8 and $g_c^{(i)}$9; a new network is initialized and trained on the union of labeled data and selected pseudo-labels; the procedure is repeated until the number of selected pseudo-labels or validation performance stabilizes. Reinitializing the network each iteration is explicitly used to limit error propagation (Rizve et al., 2021).

In DOD-SA, the PLA is inseparable from a staged teacher-student curriculum. PaST proceeds through pretraining the SM-Branch, guiding the DMD-Branch with the SM-Branch, and refining the DMD-Branch. PLA is partially active in Stage 1 through PLF, fully activated in Stage 2 through PLF+SDLM+DLC, and reused in Stage 3 with teacher pseudo labels from the DMD-Branch. The label pair memory in DLC persists across epochs rather than being recomputed statelessly (Jin et al., 14 Aug 2025).

PseCo couples PLA with Positive-proposal Consistency Voting and Multi-view Scale-invariant Learning. The teacher predicts on a weakly augmented view $y^d = \{c, b_{ir}, \tilde b_{rgb}\}$0, the student is trained on $y^d = \{c, b_{ir}, \tilde b_{rgb}\}$1 and on a second downsampled view $y^d = \{c, b_{ir}, \tilde b_{rgb}\}$2, and the same pseudo boxes supervise both views after geometric rescaling. This is noteworthy because the PLA is only one part of the unlabeled-data pathway; feature-level and label-level consistency are engineered around it (Li et al., 2022).

Point2RBox-v3 also makes assignment stage-dependent. Before the switch epoch $y^d = \{c, b_{ir}, \tilde b_{rgb}\}$3, PLA derives pseudo RBoxes from Watershed; after $y^d = \{c, b_{ir}, \tilde b_{rgb}\}$4, it trusts score-selected predictions nearest to each annotated point across FPN levels. The purpose is to restore classical multi-level assignment gradually, only after predictions become sufficiently informative (Zhang et al., 30 Sep 2025).

A recurring pattern is that PLA quality is rarely independent of the surrounding training loop. Calibration, EMA teachers, refitting, stage scheduling, class balancing, and memory mechanisms all alter the effective reliability of the assignment.

6. Empirical behavior, recurring limitations, and common misconceptions

Empirically, PLA improvements are substantial but conditional. BPLS outperforms traditional PLS methods in high-dimensional, overfitting-prone regimes for GLMs and GAMs, including EEG and sonar, yet when initial generalization is already strong, probability score dominates and BPLS may be slightly worse, as reported for banknote and breast cancer. The same paper also reports that with informative priors, iBPLS consistently outperforms non-Bayesian PLS, whereas the likelihood-only max-max criterion stays closest to the supervised baseline (Rodemann et al., 2023).

UPS reports strong results across modalities and label structures: on CIFAR-10 with 1000 labels, UPS error is $y^d = \{c, b_{ir}, \tilde b_{rgb}\}$5; on CIFAR-10 with 4000 labels, $y^d = \{c, b_{ir}, \tilde b_{rgb}\}$6; on UCF-101 with $y^d = \{c, b_{ir}, \tilde b_{rgb}\}$7 labeled data, accuracy is $y^d = \{c, b_{ir}, \tilde b_{rgb}\}$8; and on Pascal VOC2007 with $y^d = \{c, b_{ir}, \tilde b_{rgb}\}$9 labeled data, mAP is $$p(y_* \mid x_*, D) = \int p(y_* \mid x_*, \theta)\, p(\theta \mid D)\, d\theta,$$0. The paper attributes these gains to combining confidence thresholds, uncertainty thresholds, temperature scaling, and negative learning rather than to confidence filtering alone (Rizve et al., 2021).

Detection-oriented PLAs show similarly targeted gains. PseCo reports $$p(y_* \mid x_*, D) = \int p(y_* \mid x_*, \theta)\, p(\theta \mid D)\, d\theta,$$1 mAP at $$p(y_* \mid x_*, D) = \int p(y_* \mid x_*, \theta)\, p(\theta \mid D)\, d\theta,$$2 labels and $$p(y_* \mid x_*, D) = \int p(y_* \mid x_*, \theta)\, p(\theta \mid D)\, d\theta,$$3 mAP at $$p(y_* \mid x_*, D) = \int p(y_* \mid x_*, \theta)\, p(\theta \mid D)\, d\theta,$$4 labels on COCO, outperforming previous SSOD baselines while improving learning efficiency. Efficient Teacher reports $$p(y_* \mid x_*, D) = \int p(y_* \mid x_*, \theta)\, p(\theta \mid D)\, d\theta,$$5 AP$$p(y_* \mid x_*, D) = \int p(y_* \mid x_*, \theta)\, p(\theta \mid D)\, d\theta,$$6 on COCO-standard with the ResNet-50-FPN Dense Detector versus $$p(y_* \mid x_*, D) = \int p(y_* \mid x_*, \theta)\, p(\theta \mid D)\, d\theta,$$7 for supervised training, and up to $$p(y_* \mid x_*, D) = \int p(y_* \mid x_*, \theta)\, p(\theta \mid D)\, d\theta,$$8 AP$$p(y_* \mid x_*, D) = \int p(y_* \mid x_*, \theta)\, p(\theta \mid D)\, d\theta,$$9 with YOLOv5l. These numbers support the claim that PLA design is especially consequential in dense detection, where assignment noise directly affects classification, regression, and objectness branches (Li et al., 2022, Xu et al., 2023).

Cross-modality and weakly supervised detection provide clearer ablations of PLA itself. In DOD-SA on DroneVehicle, the full PLA yields IR test mAP $p(D \cup (x_i,\hat y_i)\mid D).$0 and RGB test mAP $p(D \cup (x_i,\hat y_i)\mid D).$1; removing PLF, SDLM, or DLC degrades both modalities. In Point2RBox-v3, adding PLA alone raises end-to-end AP50 from $p(D \cup (x_i,\hat y_i)\mid D).$2 to $p(D \cup (x_i,\hat y_i)\mid D).$3 and two-stage AP50 from $p(D \cup (x_i,\hat y_i)\mid D).$4 to $p(D \cup (x_i,\hat y_i)\mid D).$5. These results support a narrower interpretation of PLA as a geometry- and correspondence-sensitive assignment module rather than a generic pseudo-label filter (Jin et al., 14 Aug 2025, Zhang et al., 30 Sep 2025).

Partial-label methods report that pseudo-label quality, not merely model capacity, is often the bottleneck. SARI reaches $p(D \cup (x_i,\hat y_i)\mid D).$6 on CIFAR-10 at $p(D \cup (x_i,\hat y_i)\mid D).$7, $p(D \cup (x_i,\hat y_i)\mid D).$8 on CIFAR-100 at $p(D \cup (x_i,\hat y_i)\mid D).$9, and $\pi(\theta)=p(\theta\mid D)$0 on CUB-200 NPLL at $\pi(\theta)=p(\theta\mid D)$1. Reduction-based pseudo-label generation reports $\pi(\theta)=p(\theta\mid D)$2 on CIFAR-10, $\pi(\theta)=p(\theta\mid D)$3 on CIFAR-100, and $\pi(\theta)=p(\theta\mid D)$4 on TinyImageNet. These works explicitly argue that direct self-training can be bottlenecked by overfitting to incorrect candidate labels, and that better pseudo-label assignment is the core intervention (Saravanan et al., 2024, Qiao et al., 2024).

Several limitations recur. BPLS incurs Hessian and log-determinant overhead and relies on a reasonably well-behaved posterior. DOD-SA can fail under severe parallax or extreme field-of-view differences, in which case SDLM falls back to copying infrared boxes. Efficient Teacher remains sensitive to dense small-object regimes and to objectness calibration. Point2RBox-v3 adds computational cost through stage-dependent pseudo-label generation and mask routing. Active refinement depends on the quality of the warm-up fine-labeled set. This suggests that no PLA is universally dominant; assignment quality depends on whether the main error source is overconfidence, uncertainty miscalibration, ambiguous candidates, noisy localization, or structural mismatch.

A final misconception is that pseudo labels are always hard positive class labels. The literature reviewed here includes negative pseudo-labels, soft per-class distributions, soft objectness targets, paired cross-modality boxes, and dynamic oriented boxes. The term “assigner” is therefore literal: a PLA assigns supervision in whatever form the task and learning objective require.