Papers
Topics
Authors
Recent
Search
2000 character limit reached

Class-Aware Dense Pseudo-Label Assignment

Updated 12 July 2026
  • The paper introduces a class-aware pseudo-labeling mechanism that refines dense supervision by conditioning on semantic class data to tackle imbalances and ambiguities.
  • It employs per-class ranking and thresholding strategies to filter and reweight predictions, ensuring that even underrepresented categories are reliably labeled.
  • Empirical evaluations show significant performance gains in metrics like mIoU and mAP across diverse tasks including remote sensing detection and 3D segmentation.

to=arxiv_search.search 久赢්ඩാറ്റા ็ตทรู 天天买彩票_json code: {"11query11 Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11", "11max_results11 11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11} to=arxiv_search.search 彩神争霸 北京pk赛车_json code: {"11query11 Semantic Guidance for Sparsely Annotated Remote Sensing Object Detection\" OR 11all:\11 Teacher: Dense Pseudo-Labels for Semi-supervised Object Detection\" OR 11all:\11 Matching Semi-Supervised Object Detection\" OR 11all:\11 Semi-Supervised Learning for Dense Object Detection\" OR 11all:\11 Category Learning: A Semi-Supervised Learning Method for Dense Prediction with Extremely Limited Labels\" OR 11all:\11 Pseudo Labeling for Semi-Supervised Multi-Label Learning\" OR 11all:\11 Weakly Supervised Semantic Segmentation via Class-Aware and Geometry-Guided Pseudo-Label Refinement\"", "11max_results11 11max_results11query11} Class-aware dense pseudo-label assignment denotes a family of pseudo-labeling procedures that assign supervisory signals at dense spatial granularity—points, pixels, anchors, proposals, or class coordinates—while conditioning selection, masking, or propagation on semantic class information. In the works surveyed here, this design is motivated by recurrent failure modes of ordinary confidence-based pseudo-labeling: extreme class imbalance, sparse human annotation, selection ambiguity, assignment ambiguity, and confirmation bias. The resulting methods appear in sparsely annotated remote-sensing object detection, 11query11D weakly supervised semantic segmentation, semi-supervised object detection, semi-supervised multi-label learning, and dense prediction under extremely limited labels (&&&11query11&&&, &&&11all:\11&&&, &&&11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11&&&, &&&11max_results11&&&, &&&11query11&&&, &&&11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11&&&, &&&11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11&&&).

The common setting is dense prediction with incomplete supervision. In 11query11D weakly supervised semantic segmentation, sparse or low-cost annotated data are used instead of dense point-wise annotations, but the low quality of pseudo-labels and the insufficient exploitation of 11query11D geometric priors create significant technical bottlenecks. In sparsely annotated remote-sensing object detection, each training image may have only a fraction of its true objects labeled, and dense object distributions together with category imbalances create a mismatch between ordinary confidence-based pseudo-labeling and the real semantic frequencies. In semi-supervised object detection, one-stage detectors are reported to suffer from both selection ambiguity, because classification scores cannot properly represent localization quality, and assignment ambiguity, because samples are matched with improper labels in pseudo-label assignment (&&&11all:\11&&&, &&&11query11&&&, &&&11query11&&&).

A central observation is that globally ranked pseudo-labels are often dominated by large or frequent categories. In the 11query11D WSSS formulation of Class-Aware Label Refinement, a naïve global top-PRESERVED_PLACEHOLDER_11query11^ scheme tends to select almost entirely from large, high-density classes such as “wall” and “floor” and starve rare classes. In remote sensing, standard Dense Pseudo-Label methods rank all spatial positions by joint confidence, then threshold or pick top-PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11^ foregrounds, which ignores the fact that high-confidence scores do not always align with underrepresented classes. In semi-supervised multi-label learning, single-threshold or instance-wise pseudo-labeling may introduce false positive labels or neglect true positive ones when the number of labels per instance is unknown (&&&11all:\11&&&, &&&11query11&&&, &&&11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11&&&).

Dense assignment is therefore not merely a matter of increasing pseudo-label quantity. In the surveyed literature, dense supervision is consistently coupled to filtering, masking, reweighting, or structured assignment. This suggests that the topic is best understood as a balance between coverage and class-conditioned reliability rather than as unconditional densification.

11max_results11. Canonical mathematical formulations

A representative formulation is the class-wise selection problem in Class-Aware Label Refinement. Let PRESERVED_PLACEHOLDER_11max_results11^ be the set of points initially assigned class PRESERVED_PLACEHOLDER_11query11, let PRESERVED_PLACEHOLDER_11all:\11^ be the confidence score, and let PRESERVED_PLACEHOLDER_11 OR all:\11^ be the retention ratio. The refinement selects the top PRESERVED_PLACEHOLDER_11 OR all:\11^ points within each class by solving

PRESERVED_PLACEHOLDER_11 OR all:\11^

The closed-form solution is to sort PRESERVED_PLACEHOLDER_11 OR all:\11^ by PRESERVED_PLACEHOLDER_11 OR all:\11^ in descending order and pick the top PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11; points with PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11^ are marked “unlabeled” PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11max_results11. The same mechanism can be written through per-class cutoffs PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11, so that points of class PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11all:\11^ with score below PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11^ are masked out (&&&11all:\11&&&).

A second canonical structure uses class-wise thresholds derived from estimated class proportions. In Class-Aware Pseudo-Labeling for semi-supervised multi-label learning, each class PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11^ has two thresholds, PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11^ and PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11. For an unlabeled sample, PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11^ yields a positive pseudo-label, PRESERVED_PLACEHOLDER_11max_results11query11^ yields a negative pseudo-label, and the intermediate band is ignored. The thresholds are selected so that the fractions of predicted positives and negatives on the unlabeled set match class-proportion estimates PRESERVED_PLACEHOLDER_11max_results11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11^ and PRESERVED_PLACEHOLDER_11max_results11max_results11^ obtained from the labeled set; in practice, this is implemented through quantiles of PRESERVED_PLACEHOLDER_11max_results11query11^ (&&&11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11&&&).

A third formulation introduces semantic gating before dense assignment. In LLM-assisted semantic guidance for remote-sensing detection, the teacher provides class logits PRESERVED_PLACEHOLDER_11max_results11all:\11^ and localization quality PRESERVED_PLACEHOLDER_11max_results11 OR all:\11, and the joint confidence is

PRESERVED_PLACEHOLDER_11max_results11 OR all:\11^

Given an image-level semantic prior PRESERVED_PLACEHOLDER_11max_results11 OR all:\11, the class-aware score is

PRESERVED_PLACEHOLDER_11max_results11 OR all:\11^

with PRESERVED_PLACEHOLDER_11max_results11 OR all:\11^ if PRESERVED_PLACEHOLDER_11query11query11^ and PRESERVED_PLACEHOLDER_11query11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11^ otherwise. Positive pseudo-labels are then built by uniting a semantic filtered pool, a global top-PRESERVED_PLACEHOLDER_11query11max_results11^ pool, class-wise top-PRESERVED_PLACEHOLDER_11query11query11^ pools, and, for sparsely labeled images, the ground-truth pixels PRESERVED_PLACEHOLDER_11query11all:\11^ (&&&11query11&&&).

In semi-supervised object detection, class-aware thresholding also appears in distribution-matching form. LabelMatch sorts teacher scores for each class and sets a per-class threshold PRESERVED_PLACEHOLDER_11query11 OR all:\11^ by indexing the sorted list at PRESERVED_PLACEHOLDER_11query11 OR all:\11, where PRESERVED_PLACEHOLDER_11query11 OR all:\11^ is the number of labeled boxes of class PRESERVED_PLACEHOLDER_11query11 OR all:\11, and PRESERVED_PLACEHOLDER_11query11 OR all:\11^ and PRESERVED_PLACEHOLDER_11all:\11query11^ are the numbers of labeled and unlabeled images. The effect is to target a pseudo-box count on unlabeled data that mirrors the average number of class-PRESERVED_PLACEHOLDER_11all:\11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11^ boxes per labeled image (&&&11max_results11&&&).

11query11. Sources of class awareness

One source is explicit semantic prior information. In the remote-sensing framework, a vision-LLM and Vicuna are queried with the prompt “Choose categories present in the image: PRESERVED_PLACEHOLDER_11all:\11max_results11. Answer in one word or short phrase.” The returned subset PRESERVED_PLACEHOLDER_11all:\11query11^ is combined with annotated classes to form PRESERVED_PLACEHOLDER_11all:\11all:\11. Any class not mentioned by the LLM nor present in sparse ground truth is suppressed during dense pseudo-label scoring (&&&11query11&&&).

A second source is empirical class distribution. LabelMatch assumes that the labeled set is an approximately unbiased random sample from the same distribution as the unlabeled set and uses the labeled foreground-foreground distribution PRESERVED_PLACEHOLDER_11all:\11 OR all:\11^ and foreground-background ratio PRESERVED_PLACEHOLDER_11all:\11 OR all:\11^ to guide unlabeled pseudo-label thresholds by minimizing the KL-divergence between labeled and pseudo-labeled distributions. CAP adopts an analogous principle for multi-label learning: the empirical labeled proportion PRESERVED_PLACEHOLDER_11all:\11 OR all:\11^ is used as an estimate of the unlabeled positive rate PRESERVED_PLACEHOLDER_11all:\11 OR all:\11, with a Hoeffding-based bound

PRESERVED_PLACEHOLDER_11all:\11 OR all:\11^

holding with probability at least PRESERVED_PLACEHOLDER_11 OR all:\11query11^ (&&&11max_results11&&&, &&&11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11&&&).

A third source is local class consistency inside the dense assignment rule itself. In CALR, every class contributes its own highest-confidence points because the global ranking is replaced with PRESERVED_PLACEHOLDER_11 OR all:\11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11^ local rankings, one per class. In Ambiguity-Resistant Semi-Supervised Learning, every location PRESERVED_PLACEHOLDER_11 OR all:\11max_results11^ carries a label PRESERVED_PLACEHOLDER_11 OR all:\11query11, and candidate locations are matched only to positives of the same class through

PRESERVED_PLACEHOLDER_11 OR all:\11all:\11^

This prevents a location from being forced to learn a box for a different object category (&&&11all:\11&&&, &&&11query11&&&).

These sources are not mutually exclusive. The literature combines class awareness from prompts, empirical label statistics, per-class rankings, and same-class matching constraints, depending on the structure of the task.

11all:\11. Densification and assignment mechanisms

Dense assignment is instantiated differently across modalities. In Dense Teacher, the teacher produces a dense tensor of class-wise probabilities PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11^ over all anchors and classes, with no NMS, no class-thresholding, and no box-to-anchor assignment on these scores. A Feature Richness Score PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11^ is used to select the top-PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11^ of spatial positions, yielding a binary mask PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11, and the dense pseudo-label at location PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11^ is PRESERVED_PLACEHOLDER_11 OR all:\11query11. The dense signal is thus preserved, but supervision is concentrated on the richest anchors (&&&11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11&&&).

In the remote-sensing CLA mechanism, densification is assembled through set unions. The semantic filtered pool is

PRESERVED_PLACEHOLDER_11 OR all:\11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11^

the confidence pool is PRESERVED_PLACEHOLDER_11 OR all:\11max_results11, the class-wise pools are PRESERVED_PLACEHOLDER_11 OR all:\11query11, and the final positives are

PRESERVED_PLACEHOLDER_11 OR all:\11all:\11^

or, for sparsely labeled images,

PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11^

All locations in PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11^ become positive pseudo-labels; the remaining locations are treated as negatives (&&&11query11&&&).

In 11query11D WSSS, dense coverage is obtained iteratively. CALR first produces a partially labeled point cloud PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11. Geometry-Aware Label Refinement then fills in additional points at the superpoint level. The Label Update strategy applies CALR only on unlabeled points at iteration PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11, merges retained old labels with newly accepted labels to form PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11, and applies GALR with overlap threshold PRESERVED_PLACEHOLDER_11 OR all:\11query11, typically PRESERVED_PLACEHOLDER_11 OR all:\11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11, to re-smooth labels at each iteration. By PRESERVED_PLACEHOLDER_11 OR all:\11max_results11^ the label coverage is nearly PRESERVED_PLACEHOLDER_11 OR all:\11query11^ and the 11query11D model is trained fully densely (&&&11all:\11&&&).

In two-stage semi-supervised detection, densification may occur at the proposal level rather than strictly at the pixel or anchor level. LabelMatch splits teacher pseudo boxes into reliable and uncertain subsets. Reliable pseudo labels use standard IoU-based assignment and support classification and regression losses, while uncertain pseudo boxes are mapped to student proposals, passed through the teacher’s RoI head, and converted into proposal-wise soft class vectors PRESERVED_PLACEHOLDER_11 OR all:\11all:\11. This proposal self-assignment replaces naïve IoU-based hard labeling for uncertain boxes (&&&11max_results11&&&).

11 OR all:\11. Uncertainty, ambiguity, and confirmation bias

A major design axis concerns the treatment of ambiguous or confusing samples. Virtual Category learning defines a sample PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11^ as confusing when its Potential Category set

PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11^

has cardinality greater than PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11. Instead of discarding such samples or forcing a concrete class label, the method appends a virtual weight PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11^ and a virtual logit PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11, assigns PRESERVED_PLACEHOLDER_11 OR all:\11query11, sets PRESERVED_PLACEHOLDER_11 OR all:\11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11^ for all PRESERVED_PLACEHOLDER_11 OR all:\11max_results11, and ignores the classes in PRESERVED_PLACEHOLDER_11 OR all:\11query11. This guarantees that for all PRESERVED_PLACEHOLDER_11 OR all:\11all:\11, PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11, avoiding misleading gradients toward any wrong class in PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11^ while still using the sample in optimization (&&&11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11&&&).

Ambiguity can also be decomposed into selection ambiguity and assignment ambiguity. ARSL introduces Joint-Confidence Estimation with

PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11^

so that pseudo-label quality depends jointly on classification and localization. It then partitions locations into negatives PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11, candidates PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11, and positives PRESERVED_PLACEHOLDER_11 OR all:\11query11^ through PRESERVED_PLACEHOLDER_11 OR all:\11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11^ and a dynamic

PRESERVED_PLACEHOLDER_11 OR all:\11max_results11^

Classification uses soft one-hot targets for PRESERVED_PLACEHOLDER_11 OR all:\11query11, while localization uses hard targets for PRESERVED_PLACEHOLDER_11 OR all:\11all:\11^ and score-weighted averaged targets for selected candidates in PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11^ (&&&11query11&&&).

Other frameworks handle uncertainty by selective masking rather than explicit ambiguity classes. Dense Teacher suppresses non-selected anchors to the zero vector after top-PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11^ region selection. CAP introduces an ignore band between PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11^ and PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11. CALR marks unselected points as PRESERVED_PLACEHOLDER_11 OR all:\11 OR all:\11. LabelMatch uses a stricter threshold PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11query11^ to separate reliable from uncertain pseudo labels. A recurring misconception is that dense pseudo-labeling means training on every predicted location. Across these methods, dense supervision is almost always mediated by class-aware selection, masking, or soft-target construction (&&&11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11&&&, &&&11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11&&&, &&&11all:\11&&&, &&&11max_results11&&&).

The empirical literature consistently reports gains from replacing global or class-agnostic selection with class-aware dense assignment. In 11query11D WSSS on ScanNet val, the global top-PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11\% baseline yields PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11max_results11^ mIoU, whereas “+ CALR only” yields PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11query11^ mIoU, a PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11all:\11. Table 11 OR all:\11^ further reports that many small classes rise from near PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11 OR all:\11^ to PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11 OR all:\11^ IoU after CALR, and the ablation identifies PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11 OR all:\11^ as the sweet spot (&&&11all:\11&&&).

In sparsely annotated remote-sensing detection on DOTA, adding CLA changes performance from PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11 OR all:\11^ mAP to PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11 OR all:\11^ mAP at PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11^ labeling, from PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11^ to PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11max_results11^ at PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11, and from PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11all:\11^ to PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11^ at PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11; the oracle GT prompt further reaches PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11, PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11, and PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11, respectively. On HRSC11max_results11query11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11, the full model with CLA achieves PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11max_results11query11^ mAP versus PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11max_results11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11^ for Dense Teacher. The ablation isolating CLA reports PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11max_results11max_results11^ mAP at PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11max_results11query11^ labels and PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11max_results11all:\11^ at PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11max_results11 OR all:\11^ over the baseline with AHR only (&&&11query11&&&).

In semi-supervised object detection, Dense Teacher reports on COCO-standard with PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11max_results11 OR all:\11^ labels that supervised only gives AP PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11max_results11 OR all:\11, Unbiased Teacher gives PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11max_results11 OR all:\11, Dense Teacher without region selection gives PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11max_results11 OR all:\11, Dense Teacher plus region division for classification gives PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11query11, and the full method gives PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11. LabelMatch on COCO-standard with PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11max_results11^ labels reports PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11query11^ mAP for supervised baseline, PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11all:\11^ for Unbiased Teacher, PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11 OR all:\11^ for Soft Teacher, and PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11 OR all:\11^ for LabelMatch; on VOC, it reports PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11 OR all:\11^ APPRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11 OR all:\11^ for supervised, PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11 OR all:\11^ for Unbiased Teacher, and PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11all:\11query11^ for LabelMatch. ARSL on COCO-standard with PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11all:\11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11^ labeled data reports PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11all:\11max_results11^ AP for the FCOS baseline, PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11all:\11query11^ with JCE alone, PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11all:\11all:\11^ with TSA without candidate mining, and PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11all:\11 OR all:\11^ with full TSA (&&&11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11&&&, &&&11max_results11&&&, &&&11query11&&&).

Comparable effects appear outside standard detection. In semi-supervised multi-label learning, CAP achieves mAP PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11all:\11 OR all:\11^ on VOC at PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11all:\11 OR all:\11, PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11all:\11 OR all:\11^ on COCO at PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11all:\11 OR all:\11, and PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11query11^ on NUS-WIDE at PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11, outperforming the listed baselines. In dense prediction under extremely limited labels, Virtual Category learning reports Pascal VOC gains from PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11max_results11^ to PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11query11^ at label ratio PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11all:\11, from PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11 OR all:\11^ to PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11 OR all:\11^ at PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11 OR all:\11, and from PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11 OR all:\11^ to PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11 OR all:\11^ at PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11query11; Table VIII additionally shows that VC works with both CE-loss and MSE-loss. These results support the view that class-aware dense pseudo-label assignment is not a single algorithm but a design principle for converting partial, noisy, or ambiguous predictions into denser supervision without collapsing rare categories or amplifying confirmation bias (&&&11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11query11&&&, &&&11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11&&&).

An adjacent precursor appears in self-supervised dense descriptors. CadON assigns each sequence either a hard pseudo-class label PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11^ or a soft confidence PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11max_results11^ derived from a similarity graph, and injects these into dense descriptor learning. On cluttered test images, DON+Soft retains correct matches for PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11query11^ of points versus approximately PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11all:\11^ for the original DON, and the reported Franka Panda grasping experiment attains a PRESERVED_PLACEHOLDER_11Class-Aware Dense Pseudo-Label Assignment remote sensing object detection (Liao et al., 21 Sep 2025) Dense Teacher (Zhou et al., 2022) LabelMatch (Chen et al., 2022) ARSL (Liu et al., 2023) weakly supervised semantic segmentation (Xu et al., 17 Oct 2025)11 OR all:\11 OR all:\11^ success rate. Although this setting is not framed as semi-supervised pseudo-labeling for detection or segmentation, it shows that class-aware dense supervision can also emerge from fully self-supervised pseudo-class generation (&&&11all:\11all:\11&&&).

Topic to Video (Beta)

No one has generated a video about this topic yet.

Sign Up to Generate All Videos Subscribe on YouTube

Whiteboard

No one has generated a whiteboard explanation for this topic yet.

Sign Up to Generate

Follow Topic

Get notified by email when new papers are published related to Class-Aware Dense Pseudo-Label Assignment.

Sign Up to Follow Topic by Email