Intra-group View Sampling in Semisupervised Segmentation

• The paper introduces IVS as a group-based consistency framework that replaces standard one-image consistency with cross-view sampling, improving mIoU from 81.1 to 82.5.
• IVS is defined on a many-to-one dataset structure where each CdZnTe sample comprises 12 RGB views sharing a single ground-truth segmentation label.
• The method randomly samples three views per group—one weak and two strong perturbations—to mitigate noise from low-contrast defect boundaries and enhance model robustness.

Intra-group View Sampling (IVS) is a group-oriented consistency construction introduced for semi-supervised semantic segmentation of CdZnTe semiconductor images in "Harnessing Group-Oriented Consistency Constraints for Semi-Supervised Semantic Segmentation in CdZnTe Semiconductors" (Li et al., 18 Aug 2025). In that formulation, a group is the set of images corresponding to one physical CdZnTe sample, a view is one RGB image of that sample captured under one lighting angle, and all views in the group share one common ground-truth segmentation label. IVS replaces standard one-image weak-to-strong consistency with cross-view consistency inside the group: during each training iteration, three views are randomly sampled from the same group, one is used as the weak branch, and the other two are used as strong branches.

1. Definition and data structure

IVS is defined against a dataset organization that the paper describes as a “many-to-one” relationship. Each CdZnTe group contains 12 RGB images and all 12 views share one common ground-truth segmentation label (Li et al., 18 Aug 2025). The dataset used to instantiate this setting is TPO (“Twelve images Plus One corresponding Label”), acquired with a single-camera, multi-light-source system under 360-degree illumination conditions.

This grouping is not incidental metadata. The paper argues that standard semi-supervised semantic segmentation assumes a “one-to-one” relationship, in which each image is independently associated with its ground truth. In CdZnTe inspection, however, annotators cross-reference multiple views because a single illumination condition often provides incomplete boundary evidence. IVS therefore redefines the unlabeled consistency signal around the group rather than the individual image.

Formally, the unlabeled and labeled sets are organized as

$$\{X_{i,k}^{u}\}_{1 \leq i \leq N, 1 \leq k \leq 12}$$

and

$$\{X_{i,k}^{l}, Y_{i}^l\}_{1 \leq i \leq M, 1 \leq k \leq 12},$$

where $i$ is the group index, $k$ is the view index, and $Y_i^l$ is the shared mask for group $i$ (Li et al., 18 Aug 2025).

2. Motivation: group consistency as natural perturbation

The immediate motivation for IVS is the low-contrast nature of CdZnTe defect boundaries. The paper states that boundaries may be ambiguous in one view and clearer in another, so ordinary weak-to-strong consistency can reinforce incorrect pseudo-labels in uncertain regions (Li et al., 18 Aug 2025). The resulting failure mode is a form of confirmation bias: one view yields an incomplete pseudo-label, and consistency regularization then stabilizes that same mistake.

IVS is built on the hypothesis that views inside a CdZnTe group provide inherent consistency constraints. The paper attributes this to four properties: shared semantic content, complementary boundary clues, cross-view invariance, and a natural perturbation structure induced by changes in illumination angle. Different views of the same sample are thus treated as semantically preserving perturbations rather than unrelated images.

This design is explicitly human-inspired. Annotators do not trust one low-contrast image in isolation; they inspect several views before deciding on the defect boundary. IVS translates that practice into a training rule: one sampled view provides the reference prediction, and other sampled views from the same group are trained to agree with it (Li et al., 18 Aug 2025).

3. Mathematical formulation and sampling procedure

The baseline semi-supervised objective used for comparison is the standard image-wise weak-to-strong consistency loss: $\mathcal{L}_{con} = \frac{1}{|\mathcal{B}|}\!\sum_{i=1}^{|\mathcal{B}|}\!\frac{1}{H\times W}\sum_{j=1}^{H\!\times\!W}\!\ell(p^s_{i(j)}, p^w_{i(j)}) \mathbbm{1}(\max(p^w_{i(j)}) \geq \tau).$ Here $\ell$ is cross-entropy, $p^w$ is the weak prediction, $p^s$ is the strong prediction, and $$\{X_{i,k}^{l}, Y_{i}^l\}_{1 \leq i \leq M, 1 \leq k \leq 12},$$0 is the confidence threshold, with default $$\{X_{i,k}^{l}, Y_{i}^l\}_{1 \leq i \leq M, 1 \leq k \leq 12},$$1 (Li et al., 18 Aug 2025).

IVS reformulates this as a group-based consistency loss. During each training iteration, the method randomly samples three views from each group to construct one weak perturbation and two strong perturbations. The resulting loss is

$$\{X_{i,k}^{l}, Y_{i}^l\}_{1 \leq i \leq M, 1 \leq k \leq 12},$$2

with

$$\{X_{i,k}^{l}, Y_{i}^l\}_{1 \leq i \leq M, 1 \leq k \leq 12},$$3

The notation has a direct operational meaning. $$\{X_{i,k}^{l}, Y_{i}^l\}_{1 \leq i \leq M, 1 \leq k \leq 12},$$4 is the weak prediction from the first sampled view of group $$\{X_{i,k}^{l}, Y_{i}^l\}_{1 \leq i \leq M, 1 \leq k \leq 12},$$5; $$\{X_{i,k}^{l}, Y_{i}^l\}_{1 \leq i \leq M, 1 \leq k \leq 12},$$6 and $$\{X_{i,k}^{l}, Y_{i}^l\}_{1 \leq i \leq M, 1 \leq k \leq 12},$$7 are strong predictions from the second and third sampled views (Li et al., 18 Aug 2025). The first sampled view serves as the pseudo-label source, and the other two views are supervised against it under the confidence mask.

The sampling rule is simple and non-adaptive. The paper does not define a learned sampling distribution, nonuniform selection probability, or geometry-aware policy. Sampling is described as random. Weak perturbations follow the usual geometric operations such as resizing, cropping, and flipping, whereas strong perturbations include color jittering and blurring (Li et al., 18 Aug 2025). Consequently, IVS is best characterized as a cross-view consistency construction rather than a view-selection algorithm.

4. Position within ICAF and empirical role

IVS is the first group-oriented component in the broader Intra-group Consistency Augmentation Framework (ICAF) (Li et al., 18 Aug 2025). In the paper’s component progression, IVS defines the Group-Baseline, which directly replaces the conventional Semi-Baseline. The segmentation backbone remains standard: experiments use DeepLabV3+ with a ResNet-101 backbone pretrained on ImageNet, output stride 16, SGD with momentum, learning rate 0.001, 80 epochs, and input resolution $$\{X_{i,k}^{l}, Y_{i}^l\}_{1 \leq i \leq M, 1 \leq k \leq 12},$$8 (Li et al., 18 Aug 2025). IVS changes the organization of unlabeled supervision rather than the segmentation architecture itself.

Within ICAF, IVS is later extended by the Pseudo-label Correction Network (PCN), which includes the View Augmentation Module (VAM) and View Correction Module (VCM). The role of these later modules is to improve the quality of the pseudo-label source that IVS initially derives from a single weak-view prediction. This suggests a layered interpretation of the framework: IVS validates the usefulness of intra-group consistency, while PCN attempts to make that consistency signal less noisy.

The empirical evidence reported for IVS itself is concise but direct. In the component analysis at the 50% partition, the Semi-Baseline reaches 81.1 mIoU and the Group-Baseline (IVS) reaches 82.5 mIoU, which the paper interprets as validation of the intra-group consistency hypothesis (Li et al., 18 Aug 2025). The full ICAF then increases this result further through additional correction and augmentation modules.

Setting	mIoU	Interpretation
Semi-Baseline	81.1	Standard image-wise consistency
Group-Baseline (IVS)	82.5	IVS alone
Full ICAF at 50%	86.3	IVS plus later modules
Full ICAF at 5‰ (2 groups)	70.6	Low-label regime

The low-label result is especially prominent in the abstract: with only 2 group-annotated data (5‰), the full method achieves 70.6% mIoU on the CdZnTe dataset (Li et al., 18 Aug 2025). That number is not an IVS-only result, but it situates IVS as the foundational step in the framework that produces it.

5. Assumptions, limitations, and scope of generalization

IVS assumes that views are correctly grouped, that they are aligned well enough to share the same segmentation label, that appearance changes are label-preserving rather than geometry-changing, and that the group contains complementary evidence rather than only noise (Li et al., 18 Aug 2025). These assumptions are plausible in the TPO acquisition setup, where images of the same sample are captured under controlled illumination changes.

The method also has clear limitations. First, it is tied to datasets with reliable group structure and shared labels; the paper presents it specifically for CdZnTe semiconductor inspection. Second, IVS alone still depends on a single weak-view pseudo-label source, which can remain noisy in ambiguous regions. This is the explicit reason the authors introduce PCN afterward. Third, more views are not necessarily better once the framework moves beyond basic IVS into multi-view fusion modules: the hyperparameter study reports that increasing sampled views can eventually hurt because blurred or low-quality views introduce noise (Li et al., 18 Aug 2025).

The paper also suggests broader applicability under conditions that mirror the CdZnTe setting: multiple observations per sample, a shared segmentation target, complementary appearance changes across views, and sufficiently aligned views. This suggests transfer to other multi-observation segmentation regimes, but such transfer is an inference rather than an experimental result in the current paper.

6. Terminological ambiguity and adjacent research

The acronym IVS is not stable across recent arXiv literature. In (Li et al., 18 Aug 2025), IVS means Intra-group View Sampling. In "Active View Selection for Scene-level Multi-view Crowd Counting and Localization with Limited Labels," the exact term is Independent View Selection (IVS), a geometry-driven greedy camera-subset method for crowd counting and localization rather than a group-consistency loss (Zhang et al., 20 Sep 2025). In "EVTP-IVS: Effective Visual Token Pruning For Unifying Instruction Visual Segmentation In Multi-Modal LLMs," IVS means Instructed Visual Segmentation, which is unrelated to group-wise sampling in the CdZnTe sense (Zhu et al., 16 Aug 2025).

Paper	Meaning of IVS or closest term	Relation
(Li et al., 18 Aug 2025)	Intra-group View Sampling	Direct use of the term
(Zhang et al., 20 Sep 2025)	Independent View Selection	Different geometry-based view-selection framework
(Zhu et al., 16 Aug 2025)	Instructed Visual Segmentation	Different acronym altogether

Several nearby papers are IVS-adjacent without using the term directly. "ICG-MVSNet" models intra-view and cross-view relationships for multi-view stereo but does not introduce an IVS module or any view-group sampling policy (Hu et al., 27 Mar 2025). "G2G" represents image groups with known intra-group geometry and applies a perceiver resampler, but it has no explicit frame/view subset selection (Wei et al., 6 Jun 2026). "GIIM" constructs explicit graph edges among views of the same lesion and among lesions in the same view, yet it likewise uses all available views rather than sampling them (Sam et al., 10 Mar 2026). "IntelliCap" is a practical reference for IVS-like capture guidance because it combines scene-level coverage with object-centric angular densification through spherical proxies, but it does not formalize an IVS objective (Yasunaga et al., 18 Aug 2025). "MotionGRPO" addresses low intra-group diversity in GRPO groups through stochastic sampling and Perlin-noise perturbation, but its “views” are multiple sampled motion candidates rather than camera views (Yao et al., 7 May 2026).

A common misconception is therefore to treat IVS as a universally shared technical term. Current usage is domain-specific. In the strict sense established by (Li et al., 18 Aug 2025), IVS denotes a training-time mechanism that uses multiple views within a group as semantically preserving perturbations for semi-supervised segmentation. It is neither a generic camera-view selector nor a synonym for instructed visual segmentation.