Using In-Context Learning for Automatic Defect Labelling of Display Manufacturing Data

Babar Hussain; Bihai She; Dahai Yu; Gang Chen; Qiang Liu

REVIEW 5 major objections 5 minor 10 references

Using In-Context Learning for Automatic Defect Labelling of Display Manufacturing Data

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Auto-labelled defect masks from an in-context learning model train a display-inspection detector as well as human-annotated masks do.

desk verdict Useful applied paper on SegGPT for display defect auto-labeling, but the headline equivalence result may only hold on the easy 60% of images. read the letter →

arxiv 2506.04717 v1 pith:WWKQAA62 submitted 2025-06-05 cs.CV cs.AIcs.LG

Babar Hussain , Qiang Liu , Gang Chen , Bihai She , Dahai Yu This is my paper

classification cs.CVcs.AIcs.LG

keywords displaypanelinspectiondefectdetectionauto-labellingdeeplearningin-contextsegmentationscribbleannotationSegGPT

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that an in-context-learning segmentation model can take over a substantial share of manual labelling in display-panel defect inspection. The authors adapt the SegGPT architecture to panel-defect images, add a scribble-based prompt that lets an annotator mark a defect with a rough stroke instead of a precise mask, and train in two stages on 200,000 then 10,000 labelled circuit images. On three product types, the auto-labelling model beats the pre-trained baseline by an average of 0.22 IoU and 14% recall, and it produces acceptable masks for about 60% of unseen images. In a downstream comparison, a model trained on auto-generated labels performs about as well as one trained on human labels (IoU 0.84 vs 0.85; recall 99.16% vs 98.83%). If the result holds, manufacturers can cut annotation effort while keeping defect-detection accuracy.

What carries the argument

The load-bearing mechanism is an in-context-learning segmentation model built on SegGPT, which turns a reference image plus its mask and a query image into a mask-prediction task. Three modifications carry the argument: adaptive cropping and spatial ensembling place multiple zoomed defects into composite grids up to 4x4 so small defects remain learnable; selective label suppression randomly deactivates some defect-class labels during training so the model learns to segment only the defect class named in the prompt; and a learned scribble token lets a user prompt with a rough stroke, with paired scribble-to-mask training expanding the stroke into a full mask. The model is trained in two stages, first on 200,000 labelled LCD circuit images and then on 10,000 curated high-quality images during which the suppression and scribble mechanisms are introduced.

What would settle it

Inspect the Figure 6 workflow and re-run the downstream comparison with all auto-labelled images, including those with IoU at or below 0.60 or Hausdorff distance at or above 10 pixels; if including rejected images degrades downstream performance, the claim that auto-labelled data match human labels holds only for the accepted subset.

Watch

Extended reading notes

Core claim

The central claim is that in-context learning, in the form of an enhanced SegGPT model, can label industrial display-panel defects accurately enough to train downstream detectors as well as human annotation does. The model takes a reference image with its mask, or a scribble, as a prompt and produces a segmentation mask for a new query image. The paper reports that this auto-labelling pipeline reaches about 60% coverage—defined as predicted masks with IoU above 0.60 and Hausdorff distance below 10 pixels—across three LCD panel types, and that a detector trained on the auto-labelled data matches one trained on human labels (IoU 0.84 versus 0.85, recall 99.16% versus 98.83%). The authors present this as evidence that AI-assisted labelling is a practical way to reduce manual annotation in industrial inspection rather than a full replacement for human review.

Load-bearing premise

The downstream equivalence result assumes that the auto-labelled training data were not limited to the roughly 60% of images whose masks passed the quality threshold; if only accepted images were used, the match with human labels does not show that rejected images can be labelled automatically.

Editorial extensions

If this is right

With an average coverage rate near 60%, the auto-labelling model can pre-label more than half of unseen display-panel images, leaving human annotators to review masks rather than draw them from scratch.
A detector trained on auto-labelled data reaches IoU 0.84 and recall 99.16%, essentially matching one trained on human labels (IoU 0.85, recall 98.83%).
The reported gains of +0.22 IoU and +14% recall over the pre-trained baseline indicate that domain-specific training techniques, not the base model alone, make in-context labelling usable on industrial defect data.
The scribble-based prompting mechanism lowers the skill and time needed per annotation because a rough stroke is expanded into a complete segmentation mask.
The two-stage training pipeline runs on four consumer-grade RTX 3090 GPUs, so the approach does not depend on specialized industrial hardware.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A testable extension would be to measure end-to-end annotator time: if reviewing an AI-generated mask is faster than drawing a full mask, the roughly 40% of images below the coverage threshold still yield savings even before re-prompting.
The quality threshold (IoU above 0.60 and Hausdorff distance below 10 pixels) could be used as an explicit acquisition rule: accept auto-labels above threshold, re-prompt below it, and grow the training set iteratively.
The same scribble-token and label-suppression mechanisms could transfer to other defect-imaging domains with small, elongated, or low-contrast flaws, such as semiconductor wafers or metal surfaces, where full-mask annotation is equally costly.
A useful stress test would be to measure coverage on a product type excluded from training; the paper reports results across three product categories but does not state whether any of them were held out during fine-tuning.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a competent industry paper with real engineering value, but the central claim—auto-labels match human labels for downstream training—is under-supported as written. The most likely issue is that the comparison in Table 1 may only use the ~60% of images that passed the coverage threshold, not the full set. That needs to be clarified before the headline claim holds.

What's new: the authors adapt SegGPT to display panel defect labeling with a few sensible domain tricks: composite multi-defect augmentation with grid layouts, selective label suppression to improve prompt adherence, and a learnable scribble token for scribble-based prompting. Two-stage fine-tuning (200k images then 10k curated) on proprietary LCD data is a reasonable recipe. The reported gains over the pre-trained SegGPT baseline—0.22 average IoU and 14% recall—are plausible and consistent with the approach.

What's solid: the workflow is practical and the problem is real. The paper is honest about the 60% coverage rate and the need for human review. Using four RTX 3090s is also realistic for industrial settings.

Soft spots: (1) The Table 1 equivalence is the key result, and the paper never says whether the auto-labeled training set includes all images or only the ones that passed the coverage threshold. Figure 1's workflow explicitly returns rejected labels to the unlabeled pool, so the natural reading is that the auto-labeled stream uses only the ~60% that passed. If so, Table 1 just says the model is good on the easy subset. The stress-test note is right to flag this. (2) No error bars, no repeated runs, no per-product breakdown. A 0.01 IoU gap and 0.33% recall gap are within noise, but we can't tell. (3) The scribble token is never quantitatively evaluated. We don't learn how much accuracy drops when prompts are scribbles rather than full masks. (4) No ablation isolates the proposed components (composite augmentation, label suppression, scribble token). The baseline comparison to SegGPT is fine, but the contributions are bundled.

Citation pattern looks fine; SegGPT and Painter are properly credited, and the display-manufacturing references are appropriate.

Bottom line: this is a paper worth reviewing seriously, but it needs a revision that addresses the training-set composition question, adds error bars or repeated trials, and reports scribble and ablation results. If the Table 1 ambiguity resolves in the authors' favor, the practical claim holds. For now, I'd treat the equivalence result as unproven.

Referee Report

5 major / 5 minor

Summary. The paper presents an AI-assisted auto-labeling system for display panel defect segmentation based on an enhanced SegGPT model. The authors propose a two-stage training pipeline with adaptive cropping/spatial ensembling, multiclass training with dynamic color assignment, selective label suppression, and a scribble-based prompting mechanism using a new learnable token. They report that the enhanced model improves IoU by 0.22 and recall by 14% over the pretrained SegGPT baseline across three LCD product types, with about 60% average auto-labeling coverage (IoU>0.60, HD<10). A downstream experiment compares models trained on human-labeled versus auto-labeled data, claiming equivalent performance (IoU 0.85 vs 0.84, recall 98.83% vs 99.16%). The paper concludes that the system can reduce manual annotation effort while maintaining accuracy in industrial inspection.

Significance. If the claimed equivalence holds, this is a practically valuable contribution to reducing annotation cost in display manufacturing, an industrial domain where labeled data are scarce and expensive. The paper's strengths include a real industrial dataset, a two-stage training recipe that is plausible and reproducible with modest hardware, a comparison against the publicly available SegGPT baseline, and a downstream-task evaluation that goes beyond pixel-level metrics. The scribble-based prompting idea is interesting and could ease annotation further. However, the empirical support is currently under-developed: all metrics are point estimates without variance, the proposed components are not ablated, the scribble mechanism is not quantitatively evaluated, and the selection bias in the downstream comparison is unresolved. As presented, the central claim is plausible but not yet established to the standard expected for a journal publication.

major comments (5)

[Section 3 (Experiments and Evaluation), Figure 6, Table 1] The paper does not state whether the 'Auto Labelled Images' branch in Figure 6 consists of all images or only the subset that passed the coverage threshold (IoU>0.60, HD<10). Since Section 2 states that rejected labels are returned to the unlabeled pool and Figure 5 reports about 60% average coverage, it is likely that the auto-labeled training set used for Table 1 contains only the accepted easy subset. If so, Table 1 demonstrates equivalence only on that subset and does not establish that auto-labels are usable on the full data distribution, including the roughly 40% of images that would require re-prompting or manual labeling. Please clarify the composition of the auto-labeled training set and, if only the accepted subset was used, evaluate the downstream model on the full distribution or at least on the rejected images to quantify the gap.
[Section 3 (Experiments and Evaluation), Figures 4–5, Table 1] All reported metrics are point estimates with no error bars, confidence intervals, or repeated-run variance. The downstream differences in Table 1 (IoU 0.85 vs 0.84, recall 98.83% vs 99.16%) are small and could easily arise from training stochasticity; without variance information the claimed equivalence is not statistically supported. Please report per-run results, standard deviations or confidence intervals, and per-product-type breakdowns for Table 1, since the abstract and text claim consistency across multiple product types.
[Section 2 (Scribble-Based Annotation), Figure 3] The scribble token is a central novelty, but no experiment evaluates the scribble-to-mask path. The paper claims that scribble-based input 'substantially reduces the annotation burden while maintaining segmentation accuracy,' yet it never compares segmentation quality with scribble prompts versus full-mask prompts, nor does it report annotation time or user effort. An ablation comparing the same model with scribble and full-mask prompts, including IoU/recall differences and manual effort measurements, is needed to support this claim.
[Section 2 (Enhancing Defect Representation, Selective Label Suppression, Scribble-Based Annotation)] The proposed contributions—adaptive cropping/spatial ensembling, selective label suppression, and the scribble token—are only evaluated as a combined system. There is no ablation study that isolates any of these components. In particular, the claim that selective label suppression improves prompt faithfulness is qualitative and unsupported. Please provide an ablation that removes or toggles each proposed component and reports the resulting IoU, recall, and coverage on the same evaluation sets.
[Section 3 (Model Training)] The stage-2 curation 'filtered out extremely small defects that fell below our minimum pixel threshold,' but the threshold value is not reported, and the impact of this filtering on the evaluation is not analyzed. If the test set also contains such small defects, the reported coverage and IoU may be optimistic; if the test set excludes them, the claims do not transfer to raw inspection data. Please report the threshold, the fraction of defects removed, and the sensitivity of the results to this curation choice.

minor comments (5)

[Equation (1)] The Hausdorff distance formula is typeset with garbled subscripts and superscripts; please rewrite it cleanly and define h(A,B) and H(A,B) precisely.
[Introduction, Section 4] The Introduction states that conclusions and future work appear in Section 5, but the Conclusion is Section 4 in the provided manuscript; fix the cross-reference.
[Section 3 (Experiments and Evaluation), Figure 6] The downstream 'SegModel' used for the Table 1 comparison is not specified. Please provide the architecture, training hyperparameters, and the number of training images used for both the human-labeled and auto-labeled branches.
[Section 3 (Experiments and Evaluation), Figure 4] The reported average IoU increase of 0.22 and recall increase of 14% are stated in the abstract but not tied to a specific table or figure value; ensure the text clearly explains how these numbers are derived from Figure 4.
[Section 3 (Experiments and Evaluation), Figure 5] The coverage rate definition (IoU>0.60 and HD<10) is introduced with no sensitivity analysis or justification for these thresholds. Please report how coverage varies with threshold choices, and note any resulting impact on the practical workflow.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical comparisons against an external baseline and a downstream task, with no equation-level quantity reducing to a fitted input or self-citation chain.

full rationale

The paper's central claims are (1) that an enhanced SegGPT model improves IoU and recall over the pre-trained baseline, and (2) that a downstream model trained on auto-labeled data performs comparably to one trained on human-labeled data. Neither claim is circular in the sense defined here. The baseline is the externally published SegGPT model, and the downstream evaluation uses a separately trained segmentation model on a common test set. No equation in the paper defines an output quantity in terms of the very quantity it is supposed to predict; the coverage rate (IoU > 0.60 and HD < 10) is an evaluation metric, not a fitted parameter that is later relabeled as a prediction. The paper contains no self-citation chain: all cited prior work (SegGPT, Painter, masked autoencoders, etc.) is external. The scribble token, selective label suppression, and two-stage training are architectural/training choices evaluated empirically, not assumptions that presuppose the target result. The skeptical concern that Table 1 may compare only the roughly 60% of images that passed the coverage threshold, and that the paper does not state how rejected auto-labels were handled in the Figure 6 workflow, is a legitimate validity/selection concern about the strength of the downstream equivalence claim, but it is not circular reasoning: it does not involve an equation reducing to its own input or a load-bearing self-citation. A non-finding is therefore appropriate.

Assumptions & free parameters 2 free parameters · 4 assumptions · 1 invented entities

The central result depends on domain assumptions about defect consistency, the representativeness of a single scribbled prompt, human label accuracy, and the transferability of SegGPT fine-tuning. The coverage threshold is a hand-chosen evaluation criterion, and the scribble token is the only newly introduced architectural entity.

free parameters (2)

Coverage quality thresholds = IoU > 0.60 and HD < 10
Chosen by hand to define the 60 percent coverage rate; a different threshold would change the headline coverage number.
Minimum defect pixel threshold for stage-2 filtering = not specified
The paper filters out extremely small defects during curation without reporting the threshold value or its effect on the reported metrics.

assumptions (4)

domain assumption A single prompt image with a scribble sufficiently represents the target defect class across unseen images.
The entire auto-labeling flow relies on one-shot prompting; no analysis of intra-class variation or prompt sensitivity is provided.
ad hoc to paper The scribble-to-mask mapping learned from paired scribble and mask training transfers to real human scribbles at inference.
Section 2 introduces the scribble token but no scribble-specific evaluation is reported.
domain assumption Human ground-truth labels are accurate enough to serve as the reference for both training and evaluation.
No inter-annotator agreement or label-quality analysis is provided.
domain assumption Fine-tuning on 200,000 plus 10,000 display images preserves SegGPT's in-context generalization while adapting it to industrial defects.
Section 3 describes the two-stage procedure; there is no ablation isolating the contributions of each stage.

invented entities (1)

Scribble token
purpose: Task-based special token that distinguishes scribble-prompt conditioning from full-mask-prompt conditioning in the SegGPT encoder.
Introduced in Figure 3; the paper asserts it enables scribble-based annotation but does not provide an external benchmark or ablation isolating its effect.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Using In-Context Learning for Automatic Defect Labelling of Display Manufacturing Data." pith.science (2026). https://pith.science/paper/WWKQAA62

@misc{pith2026250604717,
  author       = {Pith},
  title        = {Pith review of: Using In-Context Learning for Automatic Defect Labelling of Display Manufacturing Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WWKQAA62}},
  note         = {Machine review of arXiv:2506.04717}
}

read the original abstract

This paper presents an AI-assisted auto-labeling system for display panel defect detection that leverages in-context learning capabilities. We adopt and enhance the SegGPT architecture with several domain-specific training techniques and introduce a scribble-based annotation mechanism to streamline the labeling process. Our two-stage training approach, validated on industrial display panel datasets, demonstrates significant improvements over the baseline model, achieving an average IoU increase of 0.22 and a 14% improvement in recall across multiple product types, while maintaining approximately 60% auto-labeling coverage. Experimental results show that models trained on our auto-labeled data match the performance of those trained on human-labeled data, offering a practical solution for reducing manual annotation efforts in industrial inspection systems.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

10 extracted references · 10 canonical work pages

[1]

32‐2: Multi AI Approaches for Improving OLED Display Pattern Repair in Manufacturing Processes

D Lim HB, Shin EC. 32‐2: Multi AI Approaches for Improving OLED Display Pattern Repair in Manufacturing Processes. InSID Symposium Digest of Technical Papers 2024 Jun (Vol. 55, No. 1, pp. 413-416)

work page 2024
[2]

A more reliable defect detection and performance improvement method for panel inspection based on artificial intelligence

Jeong EY, Kim J, Jang WH, Lim HC, Noh H, Choi JM. A more reliable defect detection and performance improvement method for panel inspection based on artificial intelligence. Journal of Information Display. 2021 Jul 3;22(3):127-36

work page 2021
[3]

P‐236: Late‐News Poster: Hard defect detection and classification for display panel products

Cho S, Kim S, Song M, Park C, Son S. P‐236: Late‐News Poster: Hard defect detection and classification for display panel products. InSID Symposium Digest of Technical Papers 2024 Jun (Vol. 55, No. 1, pp. 1617-1618)

work page 2024
[4]

81‐1: Invited Paper: Data Augmentation for Applying Deep Learning to Display Manufacturing Defect Detection

Xiong W, Lee J, Qu S, Jang W. 81‐1: Invited Paper: Data Augmentation for Applying Deep Learning to Display Manufacturing Defect Detection. InSID Symposium Digest of Technical Papers 2020 Aug (Vol. 51, No. 1, pp. 1210-1213)

work page 2020
[5]

P‐234: Late‐News Poster: A Data‐Centric Approach to Minimize Defect Leakage in an AI‐based Automated Surface Inspection System for Display Manufacturing Process

Kim SG, Jo S, Kim H, Yoo D. P‐234: Late‐News Poster: A Data‐Centric Approach to Minimize Defect Leakage in an AI‐based Automated Surface Inspection System for Display Manufacturing Process. InSID Symposium Digest of Technical Papers 2024 Jun (Vol. 55, No. 1, pp. 1609-1612)

work page 2024
[6]

Images speak in images: A generalist painter for in-context visual learning

Wang X, Wang W, Cao Y, Shen C, Huang T. Images speak in images: A generalist painter for in-context visual learning. Proc IEEE/CVF Conf Comput Vis Pattern Recognit. 2023:6830-9

work page 2023
[7]

Seggpt: Towards segmenting everything in context

Wang X, Zhang X, Cao Y, Wang W, Shen C, Huang T. Seggpt: Towards segmenting everything in context. Proc IEEE/CVF Int Conf Comput Vis. 2023:1130-40

work page 2023
[8]

Masked autoencoders are scalable vision learners

He K, Chen X, Xie S, Li Y, Dollár P, Girshick R. Masked autoencoders are scalable vision learners. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition 2022 (pp. 16000-16009)

work page 2022

Show all 10 references

[9]

Segment everything everywhere all at once

Zou X, Yang J, Zhang H, Li F, Li L, Wang J, et al. Segment everything everywhere all at once. Advances in Neural Information Processing Systems. 2024 Feb 13;36

2024
[10]

Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool

Taha AA, Hanbury A. Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool. BMC medical imaging. 2015 Dec;15:1-28. Figure

2015

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

32‐2: Multi AI Approaches for Improving OLED Display Pattern Repair in Manufacturing Processes

D Lim HB, Shin EC. 32‐2: Multi AI Approaches for Improving OLED Display Pattern Repair in Manufacturing Processes. InSID Symposium Digest of Technical Papers 2024 Jun (Vol. 55, No. 1, pp. 413-416)

work page 2024

[2] [2]

A more reliable defect detection and performance improvement method for panel inspection based on artificial intelligence

Jeong EY, Kim J, Jang WH, Lim HC, Noh H, Choi JM. A more reliable defect detection and performance improvement method for panel inspection based on artificial intelligence. Journal of Information Display. 2021 Jul 3;22(3):127-36

work page 2021

[3] [3]

P‐236: Late‐News Poster: Hard defect detection and classification for display panel products

Cho S, Kim S, Song M, Park C, Son S. P‐236: Late‐News Poster: Hard defect detection and classification for display panel products. InSID Symposium Digest of Technical Papers 2024 Jun (Vol. 55, No. 1, pp. 1617-1618)

work page 2024

[4] [4]

81‐1: Invited Paper: Data Augmentation for Applying Deep Learning to Display Manufacturing Defect Detection

Xiong W, Lee J, Qu S, Jang W. 81‐1: Invited Paper: Data Augmentation for Applying Deep Learning to Display Manufacturing Defect Detection. InSID Symposium Digest of Technical Papers 2020 Aug (Vol. 51, No. 1, pp. 1210-1213)

work page 2020

[5] [5]

P‐234: Late‐News Poster: A Data‐Centric Approach to Minimize Defect Leakage in an AI‐based Automated Surface Inspection System for Display Manufacturing Process

Kim SG, Jo S, Kim H, Yoo D. P‐234: Late‐News Poster: A Data‐Centric Approach to Minimize Defect Leakage in an AI‐based Automated Surface Inspection System for Display Manufacturing Process. InSID Symposium Digest of Technical Papers 2024 Jun (Vol. 55, No. 1, pp. 1609-1612)

work page 2024

[6] [6]

Images speak in images: A generalist painter for in-context visual learning

Wang X, Wang W, Cao Y, Shen C, Huang T. Images speak in images: A generalist painter for in-context visual learning. Proc IEEE/CVF Conf Comput Vis Pattern Recognit. 2023:6830-9

work page 2023

[7] [7]

Seggpt: Towards segmenting everything in context

Wang X, Zhang X, Cao Y, Wang W, Shen C, Huang T. Seggpt: Towards segmenting everything in context. Proc IEEE/CVF Int Conf Comput Vis. 2023:1130-40

work page 2023

[8] [8]

Masked autoencoders are scalable vision learners

He K, Chen X, Xie S, Li Y, Dollár P, Girshick R. Masked autoencoders are scalable vision learners. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition 2022 (pp. 16000-16009)

work page 2022