Pith. sign in

REVIEW 5 major objections 5 minor 78 references

A 9.9-million-parameter RGB-D network claims state-of-the-art rail defect detection, running at 162 fps.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 10:40 UTC pith:UJP5FZ6L

load-bearing objection The architecture is a plausible recombination of known blocks, but the paper's own tables contradict its headline SOTA claim, and the citation list is padded with irrelevant entries. the 5 major comments →

arxiv 2601.09118 v2 pith:UJP5FZ6L submitted 2026-01-14 cs.CV

LPCAN: Lightweight Pyramid Cross-Attention Network for Rail Surface Defect Detection Using RGB-D Data

Jackie Alex , Guoqiang Huan This is my paper
classification cs.CV
keywords rail defect detectionRGB-D fusionlightweight networkcross-attention modulespatial feature extractorsalient object detectionMobileNetv2real-time inspection
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proposes LPCANet, a lightweight dual-stream network that fuses color and depth images to detect surface defects on rails. It claims that LPCANet outperforms 18 existing methods on three RGB-D rail datasets while using only 9.90 million parameters, 2.50 GFLOPs, and 37.95 MB of storage. The core idea is that a carefully designed lightweight fusion—MobileNetv2 for RGB, a small pyramid module for depth, cross-attention to combine them, and a spatial feature extractor that emphasizes horizontal and vertical contours—can deliver both accuracy and real-time speed. If true, this matters because it would bring high-accuracy defect inspection closer to practical deployment on resource-constrained railway inspection systems. The paper also reports the model generalizes to non-rail defect datasets, suggesting the architecture captures broadly useful industrial surface-anomaly patterns.

Core claim

The paper introduces LPCANet, a lightweight pyramid cross-attention network for RGB-D rail surface defect detection. The architecture uses MobileNetv2 to extract multi-scale RGB features, a lightweight pyramid module (LPM) to generate matching depth features, a cross-attention module (CAM) that treats RGB features as queries and depth features as keys/values to fuse the two modalities, and a spatial feature extractor (SFE) built from 1×3 and 3×1 convolutions to capture horizontal and vertical structural contours. The authors claim this network achieves state-of-the-art results on NEU-RSDDS-AUG, RSDD-TYPE1, and RSDD-TYPE2, with a parameter count of 9.90M, computational cost of 2.50 GFLOPs, an

What carries the argument

The cross-attention module (CAM) is the central fusion mechanism: RGB features are linearly projected into queries, depth features into keys and values, then multi-head attention is computed at each matching resolution, letting color/texture and depth geometry interact without aggressive channel compression. Around it sit two supporting components: the lightweight pyramid module (LPM), which uses a 4×4 stride-4 conv and two conv-BN-ReLU layers to build depth features at four scales, and the spatial feature extractor (SFE), which applies 1×3 and 3×1 convolutions in parallel to pick out horizontal and vertical contours. A pixel-shuffle upsampling stage then restores the mask to full resolution

Load-bearing premise

The claim that LPCANet beats all 18 prior methods assumes that every method was evaluated under the same train/test protocol on all three rail datasets; the paper calls the datasets 'unsupervised' in §4.1.2 while using ground-truth labels for BCE training in §3.6, and it does not provide the exact splits, baseline training configurations, or code needed to confirm that assumption.

What would settle it

Run a controlled reproduction: fix train/test splits and preprocessing for NEU-RSDDS-AUG, RSDD-TYPE1, and RSDD-TYPE2, then train LPCANet and all 18 baseline models under identical schedules, hardware, and data augmentation. If any baseline reaches or exceeds LPCANet's reported mAP, IOU, or Sα on the same test split, the state-of-the-art claim is falsified. Additionally, verify whether the datasets provide pixel-level ground-truth masks; if they do not, the BCE loss in §3.6 cannot be applied as stated, which would undermine the entire training description.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Rail surface inspection could run in real time on a single consumer GPU or edge device, since the model reportedly processes 320×320 RGB-D frames at 162.60 fps with only 2.50 GFLOPs.
  • The architecture demonstrates that cross-attention plus a small spatial-feature extractor can replace much heavier fusion modules without sacrificing accuracy, pointing toward simpler designs for multimodal industrial inspection.
  • The reported gains over CSEPNet and other RGB-D SOD methods suggest that lightweight dual-stream networks can outperform larger models on domain-specific defect data, not just on general salient-object benchmarks.
  • The positive results on DAGM2007, MT, and Kolektor-SDD2 imply the same network may transfer to other industrial surface-inspection tasks, such as detecting defects on metals, textiles, or manufactured components.
  • Ablation findings—SFE at stages 1–3 is best, pixel shuffle beats transposed convolution, and a 64-channel LPM projection is optimal—provide direct design guidelines for engineers building lightweight RGB-D detectors.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The claimed state-of-the-art ranking hinges on every baseline being trained and tested under the same protocol on all three datasets; the paper does not provide split definitions, baseline training hyperparameters, or code, so the relative improvements should be treated as conditional until reproduced under matched conditions.
  • The datasets are described as 'unsupervised' in §4.1.2, yet the loss in §3.6 uses pixel-level ground-truth masks; this suggests the data actually come with labels for supervised training, and the 'unsupervised' wording may be a misnomer or reflect a different pretraining stage—either way, the evaluation protocol needs clarification.
  • The SFE's 1×3 and 3×1 separable-style convolutions could be acting like a cheap, local strip-wise attention mechanism; a direct comparison against standard 3×3 convolutions or full separable convolutions would reveal whether the specific factorization or simply the added capacity drives the gain.
  • If the generalization results hold under matched evaluation, the same dual-stream lightweight recipe could be applied to other depth-sensing industrial scenarios (e.g., bridge inspection, automotive parts), and further compression via pruning or distillation may push it toward edge deployment.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes LPCANet, a lightweight dual-stream network for RGB-D rail surface defect detection. The RGB stream uses MobileNetv2 and the depth stream uses a lightweight pyramid module (LPM); a cross-attention module (CAM) fuses the two modalities, and a spatial feature extractor (SFE) captures horizontal/vertical structure. The model is trained with binary cross-entropy and evaluated on three rail datasets (NEU-RSDDS-AUG, RSDD-TYPE1, RSDD-TYPE2) and three non-rail datasets (DAGM2007, MT, Kolektor-SDD2). The headline claims are state-of-the-art accuracy versus 18 methods, high efficiency (9.90 M parameters, 2.50 G FLOPs, 162.60 fps), and generalization to non-rail defect detection. The paper also reports ablations on SFE placement, upsampling, LPM channel width, and backbone choice.

Significance. If the headline claims were correct, LPCANet would be a practically valuable efficiency/accuracy advance for industrial inspection, with a compact RGB-D model beating much larger prior work. The architecture is clearly described and the ablation logic is coherent. However, the evidence in the manuscript does not support the claimed state-of-the-art status: the paper's own tables contain numbers that contradict the abstract and Section 5.3.1 on multiple metrics, and the non-rail generalization claim is directly contradicted by Table 4. The lack of code, dataset splits, and complete comparison tables prevents independent verification. The architecture may still have merit, but the central evaluation claims are not established.

major comments (5)
  1. [Abstract and §5.3.1, Table 1] The abstract claims '+1.77% in MAE over the best-performing baseline,' and §5.3.1 repeats this as an improvement over CSEPNet. Table 1, however, shows C2FNet with MAE 6.52, which is better (lower) than LPCANet's 7.11. The claim only holds if CSEPNet (MAE 8.88) is arbitrarily chosen as the 'best-performing' MAE baseline, ignoring C2FNet. Thus the state-of-the-art MAE claim is false as stated. Table 1 also shows DACNet with higher Fmax (88.68 vs 88.57) and CSEPNet with higher Emax (92.37 vs 92.17), so LPCANet is not state-of-the-art on multiple reported metrics.
  2. [§4.2.4 and Table 4] The text states that LPCANet shows 'consistent improvements in mAP, MAE, and IOU' on non-rail datasets. Table 4 contradicts this: XMSNet has higher IOU on all three datasets (DAGM2007: 88.90 vs 83.98; MT: 91.78 vs 82.71; Kolektor-SDD2: 87.65 vs 82.83), and on DAGM2007 XMSNet also has lower MAE (8.54 vs 8.71). Thus the generalization claim of consistent improvement is unsupported and, for IOU, refuted by the paper's own data.
  3. [Tables 2 and 3] The quantitative comparisons on RSDD-TYPE1 and RSDD-TYPE2 are truncated: rows are replaced by '...', showing only two baselines plus LPCANet. The accompanying text claims LPCANet 'consistently outperforms' several methods on these datasets, but the missing rows could contain higher-scoring baselines. Without complete tables, the cross-dataset state-of-the-art claim cannot be verified. This is a load-bearing omission for the central claim.
  4. [§4.1.2, §3.6, and reproducibility] The evaluation protocol is unclear. The datasets are described as 'unsupervised' in the abstract and §4.1.2, yet training uses ground-truth labels with binary cross-entropy (§3.6). No URLs, train/test splits for RSDD-TYPE1/TYPE2, baseline training details, or code are provided. It is therefore impossible to confirm that all 19 methods were evaluated under identical conditions. Reproducibility of the reported comparisons is essential for the state-of-the-art claim.
  5. [§4.3.1–§4.3.3 and test-set selection] The ablation studies select the final architecture using the test benchmark: the number of SFE stages is chosen from Table 5, the upsampling method from Figure 5, and the LPM channel width from Figure 6, all based on test-set metrics. This means the reported final test numbers are partly the result of selection on the test set and are not independent evaluations of those design choices. The claim that the selected configuration is optimal is therefore weaker than presented.
minor comments (5)
  1. [References and Related Work] Many references are unrelated to rail defect detection or RGB-D SOD (e.g., [32] Pride and Prejudice, [56] Afrikaans culture industry, [30]/[35] graph RAG). Related Work sections contain substantial padding with LLM/RAG literature that is not connected to the method. This obscures the actual positioning and related work.
  2. [Equation (3)] The symbols \hat{Q}^r_i, \hat{K}^d_i, \hat{V}^d_i are used in Eq. (3) before being defined in the following sentence. Define the reshaped quantities before presenting the equation.
  3. [Equation (10)] The pixel-shuffle operation PS(·) is said to restore resolution, but the input F_down and output F_mask resolution are not stated explicitly. Clarify whether the output is H×W or H/4×W/4.
  4. [Tables 5 and 6] Table 5 is hard to read because the checkmarks and metric columns are compressed; the first row appears to run values together. Table 6 omits Emax and Fmax for some rows. Reformat for clarity.
  5. [Figures 5–8] Several figures (5, 6, 7, 8) are mentioned and described in the text but contain no visible captions or quantitative detail in the manuscript provided. If the figures are present in the final version, ensure captions fully describe the curves and that the claims about Figure 5 (pixel shuffle advantages) and Figure 6 (channel degradation) are directly readable.

Circularity Check

1 steps flagged

Headline SOTA is partly self-referential: the final architecture (SFE stages, upsampling method, LPM channels) is selected on the same test set later reported as SOTA, making the benchmark comparison a selection artifact rather than an independent evaluation.

specific steps
  1. fitted input called prediction [Section 4.3.1 (Table 5), with supporting selections in Section 4.3.2 (Figure 5) and Section 4.3.3 (Figure 6); test split defined in Section 4.1.2]
    "The baseline configuration (SFE applied at stages 1-3) achieves optimal performance across most metrics. Applying SFE at stage 4 causes performance degradation (-0.22% mAP, -0.36% IOU) while increasing parameters (+0.83M) and FLOPs (+0.09G)."

    The ablation study selects the final architecture because it performs best on the NEU-RSDDS-AUG test set (the only evaluation split described, 362 test images), then the paper reports the same test-set numbers as LPCANet's state-of-the-art result. There is no separate validation split mentioned. Thus the final configuration is fit to the test benchmark, and the reported +1.48% S-alpha, +0.86% IOU, +1.77% MAE improvements over baselines are not an independent prediction; they are partly a maximum over configurations evaluated on that same test set. Baselines were not given the same test-set-based tuning, so the comparison is statistically forced by construction.

full rationale

The central architecture derivation (MobileNetv2 + LPM + CAM + SFE) is a standard empirical construction, not a theorem derived from premises that already contain the result. There are no load-bearing self-citations: the authors do not cite their own prior work, and no uniqueness theorem or ansatz is imported from their own publications. The main circularity is procedural: the final network configuration is selected on the test set used to report SOTA. Section 4.3.1 chooses SFE stages 1-3 because that 'achieves optimal performance' on the benchmark, and Sections 4.3.2-4.3.3 similarly pick pixel shuffle and 64 channels based on the same benchmark, after which those same numbers are presented as the model's headline gains. This makes the central SOTA claim partially self-referential. The paper also contains internal contradictions (e.g., C2FNet's MAE of 6.52 beats LPCANet's 7.11 in Table 1, and XMSNet's IOU beats LPCANet on all three non-rail datasets in Table 4), but these are correctness risks, not circularity, and are therefore not scored here. No definitional circularity, self-citation chains, or renamed known results were found.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The architecture introduces three new module names, all only internally validated. The headline efficiency numbers depend on hand-chosen hyperparameters (channel dims, SFE stages, input size, attention heads) and on an assumed benchmark protocol; no code or data are released.

free parameters (5)
  • LPM channel progression C^d = [64,128,256,512] = 64,128,256,512
    Hand-chosen scale in §3.2; ablation in Fig. 6 shows 96 degrades performance, so the number is tuned to the benchmark.
  • Number of SFE stages (3, stage-4 removed) = 3
    Chosen from Table 5 ablation on the test set; stage-1-only actually yields higher mAP (94.79), so the choice is not uniquely optimal.
  • Cross-attention heads N_h and per-head dim d_z = not reported
    Equation (3) uses N_h and d_z, but the paper never gives their values, blocking replication.
  • Input resolution = 320×320
    Set in §4.1.1; directly affects FLOPs and fps claims.
  • Training hyperparameters (AdamW LR, weight decay, momentum) = LR=1e-4, wd=0.05, momentum=0.9
    Chosen in §4.1.1; standard but underdetermine performance.
axioms (5)
  • domain assumption Dataset labels and split integrity
    The paper uses BCE ground-truth labels but calls datasets "unsupervised"; no split is given for RSDD-TYPE1/2, so label correctness and train/test separation are assumed (§4.1.2, §3.6).
  • domain assumption RGB-D pair alignment
    Cross-attention fusion assumes corresponding pixels at the same resolution; no rectification/alignment procedure is described (§3.1).
  • domain assumption Protocol parity across 19 methods
    Tables compare against 18 baseline numbers not produced here; the paper assumes identical preprocessing and splits, but baseline training details are not given (§4.2, Tables 1–4).
  • domain assumption ImageNet pretraining transfers
    MobileNetv2 pretrained on IN-1k is used to initialize the RGB branch; transferability to rail surfaces is assumed without evidence (§4.1.1).
  • standard math Standard backpropagation and optimization convergence
    Training relies on standard deep-learning optimization assumptions; not proved but routine.
invented entities (3)
  • Lightweight Pyramid Module (LPM) no independent evidence
    purpose: Extract multi-scale features from depth images.
    Effectiveness rests solely on the paper's ablations; no code or external use is provided.
  • Cross-Attention Module (CAM) no independent evidence
    purpose: Fuse RGB and depth features at corresponding resolutions.
    Reported as a key component, but only internal ablations support it; hyperparameters N_h and d_z are unreported.
  • Spatial Feature Extractor (SFE) no independent evidence
    purpose: Capture horizontal and vertical structural features from fused maps.
    No external validation or code; the stage-3 configuration is selected via test-set ablations.

pith-pipeline@v1.3.0-alltime-deepseek · 16399 in / 16821 out tokens · 138210 ms · 2026-08-03T10:40:34.151196+00:00 · methodology

0 comments
read the original abstract

This paper addresses the limitations of current vision-based rail defect detection methods, including high computational complexity, excessive parameter counts, and suboptimal accuracy. We propose a Lightweight Pyramid Cross-Attention Network (LPCANet) that leverages RGB-D data for efficient and accurate defect identification. The architecture integrates MobileNetv2 as a backbone for RGB feature extraction with a lightweight pyramid module (LPM) for depth processing, coupled with a cross-attention mechanism (CAM) for multimodal fusion and a spatial feature extractor (SFE) for enhanced structural analysis. Comprehensive evaluations on three unsupervised RGB-D rail datasets (NEU-RSDDS-AUG, RSDD-TYPE1, RSDD-TYPE2) demonstrate that LPCANet achieves state-of-the-art performance with only 9.90 million parameters, 2.50 G FLOPs, and 162.60 fps inference speed. Compared to 18 existing methods, LPCANet shows significant improvements, including +1.48\% in $S_{\alpha}$, +0.86\% in IOU, and +1.77\% in MAE over the best-performing baseline. Ablation studies confirm the critical roles of CAM and SFE, while experiments on non-rail datasets (DAGM2007, MT, Kolektor-SDD2) validate its generalization capability. The proposed framework effectively bridges traditional and deep learning approaches, offering substantial practical value for industrial defect inspection. Future work will focus on further model compression for real-time deployment.

Figures

Figures reproduced from arXiv: 2601.09118 by Guoqiang Huan, Jackie Alex.

Figure 1
Figure 1. Figure 1: Process of LPCANet model 8 [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Pixel distribution of LPCANet feature maps after channel compression ((a) stage 1; [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: SFE module architecture The convolutional operators Conv1×3 and Conv3×1 extract horizontal and vertical contour information respectively, while batch normalization (BN) and ReLU activation ensure stable training and non-linear transformation capability. 3.5 Feature Integration and Upsampling To maintain information flow throughout the network, residual connections concatenate CAM and SFE outputs before dow… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of computational load between LPCANet and state-of-the-art models [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of different sampling methods [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Influence of LPM channel dimension [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Influence of different components on segmentation mask [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of PR and ROC curves between various backbone networks, top-level [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

78 extracted references · 34 linked inside Pith

  1. [1]

    A systematic review of machine learning applications in infectious disease prediction, diagnosis, and outbreak forecasting,

    Y . Wang, J. Zhong, and R. Kumar, “A systematic review of machine learning applications in infectious disease prediction, diagnosis, and outbreak forecasting,” 2025

    2025

  2. [2]

    Real-time idling vehicles detection using combined audio- visual deep learning,

    X. Li, T. Mangin, S. Saha, R. Mohammed, E. Blanchard, D. Tang, H. Poppe, O. Choi, K. Kelly, and R. Whitaker, “Real-time idling vehicles detection using combined audio- visual deep learning,” inEmerging Cutting-Edge Developments in Intelligent Traffic and Transportation Systems. IOS Press, 2024, pp. 142–158

    2024

  3. [3]

    Joint audio-visual idling vehicle detection with streamlined input dependencies,

    X. Li, R. Mohammed, T. Mangin, S. Saha, K. Kelly, R. Whitaker, and T. Tasdizen, “Joint audio-visual idling vehicle detection with streamlined input dependencies,” inProceedings of the Winter Conference on Applications of Computer Vision, 2025, pp. 885–894

    2025

  4. [4]

    Improving text embeddings with large language models,

    L. Wang, N. Yang, X. Huang, L. Yang, R. Majumder, and F. Wei, “Improving text embeddings with large language models,”arXiv preprint arXiv:2401.00368, 2023

    Pith/arXiv arXiv 2023

  5. [5]

    Is chatgpt a good nlg evaluator? a preliminary study,

    J. Wang, Y . Liang, F. Meng, Z. Sun, H. Shi, Z. Li, J. Xu, J. Qu, and J. Zhou, “Is chatgpt a good nlg evaluator? a preliminary study,”arXiv preprint arXiv:2303.04048, 2023

    Pith/arXiv arXiv 2023

  6. [6]

    Audio and multiscale visual cues driven cross-modal transformer for idling vehicle detection,

    X. Li, R. Whitaker, and T. Tasdizen, “Audio and multiscale visual cues driven cross-modal transformer for idling vehicle detection,”arXiv preprint arXiv:2504.16102, 2025

    arXiv 2025

  7. [7]

    Setransformer: A hybrid attention-based architecture for robust human activity recognition,

    Y . Liu, X. Qin, Y . Gao, X. Li, and C. Feng, “Setransformer: A hybrid attention-based architecture for robust human activity recognition,”INNO-PRESS: Journal of Emerging Applied AI, vol. 1, no. 1, 2025

    2025

  8. [8]

    Evaluating supervised learning models for fraud detection: A comparative study of classical and deep architectures on imbalanced transaction data,

    C. Wang, C. Nie, and Y . Liu, “Evaluating supervised learning models for fraud detection: A comparative study of classical and deep architectures on imbalanced transaction data,” arXiv preprint arXiv:2505.22521, 2025

    arXiv 2025

  9. [9]

    Gated multimodal graph learning for personalized recommendation,

    S. Liu, Y . Zhang, X. Li, Y . Liu, C. Feng, and H. Yang, “Gated multimodal graph learning for personalized recommendation,”INNO-PRESS: Journal of Emerging Applied AI, vol. 1, no. 1, 2025

    2025

  10. [10]

    Extraction/conversion of geometric di- mensions and tolerances for machining features,

    J. Gao, D. T. Zheng, N. Gindy, and D. Clark, “Extraction/conversion of geometric di- mensions and tolerances for machining features,”International Journal of Advanced Manufacturing Technology, vol. 26, no. 4, pp. 405–414, Aug 2005

    2005

  11. [11]

    A datum-based model for practicing geometric dimensioning and tolerancing,

    W. Sun and Y . Gao, “A datum-based model for practicing geometric dimensioning and tolerancing,”Journal of Engineering Technology, vol. 35, pp. 38–47, Sep 2018

    2018

  12. [12]

    Tolerance information extraction for mechanical engineering drawings: A digital image processing and deep learning-based model,

    Y . Xuet al., “Tolerance information extraction for mechanical engineering drawings: A digital image processing and deep learning-based model,”CIRP Journal of Manufacturing Science and Technology, vol. 50, pp. 55–64, Jun 2024

    2024

  13. [13]

    You only look once: Unified, real-time object detection,

    J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” May 2016, preprint

    2016

  14. [14]

    tesseract-ocr/tesseract,

    tesseract-ocr, “tesseract-ocr/tesseract,” https://github.com/tesseract-ocr/tesseract, 2024, accessed: 2024-09-27. 22

    2024

  15. [15]

    Leading image & video data annotation platform CV AT,

    “Leading image & video data annotation platform CV AT,” https://www.cvat.ai, accessed: 2025-03-22

    2025

  16. [16]

    torchvision.transforms torchvision master documentation,

    “torchvision.transforms torchvision master documentation,” https://pytorch.org/vision/0.9/ transforms.html, accessed: 2025-03-22

    2025

  17. [17]

    Few could be better than all: Feature sampling and grouping for scene text detection,

    J. Tang, W. Zhang, H. Liu, M. Yang, B. Jiang, G. Hu, and X. Bai, “Few could be better than all: Feature sampling and grouping for scene text detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 4563–4572

    2022

  18. [18]

    Spts v2: single-point scene text spotting,

    Y . Liu, J. Zhang, D. Peng, M. Huang, X. Wang, J. Tang, C. Huang, D. Lin, C. Shen, X. Bai et al., “Spts v2: single-point scene text spotting,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

    2023

  19. [19]

    Integration of deep learning for automatic recognition of 2D engineering drawings,

    Y .-H. Lin, Y .-H. Ting, Y .-C. Huang, K.-L. Cheng, and W.-R. Jong, “Integration of deep learning for automatic recognition of 2D engineering drawings,”Machines, vol. 11, no. 8, Aug 2023

    2023

  20. [20]

    Fine-tuning vision-language model for automated engineering drawing information extraction,

    M. T. Khan, L. Chen, Y . H. Ng, W. Feng, N. Y . J. Tan, and S. K. Moon, “Fine-tuning vision-language model for automated engineering drawing information extraction,” 2024, preprint

    2024

  21. [21]

    AutoCAD mechanical 2022 help | about balloons (autocad mechanical toolset) | autodesk,

    “AutoCAD mechanical 2022 help | about balloons (autocad mechanical toolset) | autodesk,” https://help.autodesk.com/view/AMECH_PP/2022/ENU/?guid= GUID-F12F0EA0-0810-42EE-A3FE-327041AFAEEE, accessed: 2024-09-27

    2022

  22. [22]

    Data management and SPC software,

    “Data management and SPC software,” https://measurlink.com/, accessed: 2024-09-27

    2024

  23. [23]

    Retrieval-augmented generation for knowledge-intensive nlp tasks,

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. Küttler, M. Lewis, W.-t. Yih, T. Rocktäschelet al., “Retrieval-augmented generation for knowledge-intensive nlp tasks,”Advances in Neural Information Processing Systems, vol. 33, pp. 9459–9474, 2020

    2020

  24. [24]

    Retrieval-augmented generation for large language models: A survey,

    Y . Gao, Y . Xiong, X. Gao, K. Jia, J. Pan, Y . Bi, Y . Dai, J. Sun, and H. Wang, “Retrieval-augmented generation for large language models: A survey,”arXiv preprint arXiv:2312.10997, 2023

    Pith/arXiv arXiv 2023

  25. [25]

    Optimal boxes: boosting end- to-end scene text recognition by adjusting annotated bounding boxes via reinforcement learning,

    J. Tang, W. Qian, L. Song, X. Dong, L. Li, and X. Bai, “Optimal boxes: boosting end- to-end scene text recognition by adjusting annotated bounding boxes via reinforcement learning,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 233–248

    2022

  26. [26]

    Character recognition competition for street view shop signs,

    J. Tang, W. Du, B. Wang, W. Zhou, S. Mei, T. Xue, X. Xu, and H. Zhang, “Character recognition competition for street view shop signs,”National Science Review, vol. 10, no. 6, p. nwad141, 2023

    2023

  27. [27]

    Gpt-4 technical report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkatet al., “Gpt-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

    Pith/arXiv arXiv 2023

  28. [28]

    The llama 3 herd of models,

    A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schel- ten, A. Yang, A. Fanet al., “The llama 3 herd of models,”arXiv preprint arXiv:2407.21783, 2024. 23

    Pith/arXiv arXiv 2024

  29. [29]

    Introducing contextual retrieval,

    Anthropic, “Introducing contextual retrieval,” https://www.anthropic.com/news/ contextual-retrieval, 2024, accessed: 2024-11-02

    2024

  30. [30]

    From local to global: A graph rag approach to query-focused summarization,

    D. Edge, H. Trinh, N. Cheng, J. Bradley, A. Chao, A. Mody, S. Truitt, and J. Larson, “From local to global: A graph rag approach to query-focused summarization,”arXiv preprint arXiv:2404.16130, 2024

    Pith/arXiv arXiv 2024

  31. [31]

    Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression,

    H. Jiang, Q. Wu, X. Luo, D. Li, C.-Y . Lin, Y . Yang, and L. Qiu, “Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression,” 2024. [Online]. Available: https://arxiv.org/abs/2310.06839

    Pith/arXiv arXiv 2024

  32. [32]

    Austen,Pride and Prejudice

    J. Austen,Pride and Prejudice. Urbana, Illinois: Project Gutenberg, 2006. [Online]. Available: https://www.gutenberg.org/ebooks/1342

    2006

  33. [33]

    Translation and fusion improves zero-shot cross-lingual information extraction,

    Y . Chen, V . Shah, and A. Ritter, “Translation and fusion improves zero-shot cross-lingual information extraction,” 2024. [Online]. Available: https://arxiv.org/abs/2305.13582

    Pith/arXiv arXiv 2024

  34. [34]

    How good are llms at relation extraction under low-resource scenario? comprehensive evaluation,

    D. Jinensibieke, M. Maimaiti, W. Xiao, Y . Zheng, and X. Wang, “How good are llms at relation extraction under low-resource scenario? comprehensive evaluation,” 2024. [Online]. Available: https://arxiv.org/abs/2406.11162

    Pith/arXiv arXiv 2024

  35. [35]

    Medical graph rag: Towards safe medical large language model via graph retrieval-augmented generation,

    J. Wu, J. Zhu, Y . Qi, J. Chen, M. Xu, F. Menolascina, and V . Grau, “Medical graph rag: Towards safe medical large language model via graph retrieval-augmented generation,”

  36. [36]

    Ragas: Automated evaluation of retrieval augmented generation,

    S. Es, J. James, L. Espinosa-Anke, and S. Schockaert, “Ragas: Automated evaluation of retrieval augmented generation,” 2023. [Online]. Available: https: //arxiv.org/abs/2309.15217

    Pith/arXiv arXiv 2023

  37. [37]

    Judging llm-as-a-judge with mt-bench and chatbot arena,

    L. Zheng, W.-L. Chiang, Y . Sheng, S. Zhuang, Z. Wu, Y . Zhuang, Z. Lin, Z. Li, D. Li, E. Xinget al., “Judging llm-as-a-judge with mt-bench and chatbot arena,”Advances in Neural Information Processing Systems, vol. 36, pp. 46 595–46 623, 2023

    2023

  38. [38]

    Hermes 3 technical report,

    R. Teknium, J. Quesnelle, and C. Guang, “Hermes 3 technical report,” 2024. [Online]. Available: https://arxiv.org/abs/2408.11857

    Pith/arXiv arXiv 2024

  39. [39]

    Late chunking: con- textual chunk embeddings using long-context embedding models,

    M. Günther, I. Mohr, D. J. Williams, B. Wang, and H. Xiao, “Late chunking: con- textual chunk embeddings using long-context embedding models,”arXiv preprint arXiv:2409.04701, 2024

    Pith/arXiv arXiv 2024

  40. [40]

    Enhancing knowledge graph construction using large language models,

    M. Trajanoska, R. Stojanov, and D. Trajanov, “Enhancing knowledge graph construction using large language models,” 2023. [Online]. Available: https://arxiv.org/abs/2305.04676

    Pith/arXiv arXiv 2023

  41. [41]

    Mtvqa: Benchmarking multilingual text-centric visual question answering,

    J. Tang, Q. Liu, Y . Ye, J. Lu, S. Wei, C. Lin, W. Li, M. F. F. B. Mahmood, H. Feng, Z. Zhao et al., “Mtvqa: Benchmarking multilingual text-centric visual question answering,”arXiv preprint arXiv:2405.11985, 2024

    Pith/arXiv arXiv 2024

  42. [42]

    Multi- modal in-context learning makes an ego-evolving scene text recognizer,

    Z. Zhao, J. Tang, C. Lin, B. Wu, C. Huang, H. Liu, X. Tan, Z. Zhang, and Y . Xie, “Multi- modal in-context learning makes an ego-evolving scene text recognizer,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 15 567–15 576. 24

    2024

  43. [43]

    Mctbench: Multimodal cognition towards text-rich visual scenes benchmark,

    B. Shan, X. Fei, W. Shi, A.-L. Wang, G. Tang, L. Liao, J. Tang, X. Bai, and C. Huang, “Mctbench: Multimodal cognition towards text-rich visual scenes benchmark,”arXiv preprint arXiv:2410.11538, 2024

    Pith/arXiv arXiv 2024

  44. [44]

    Harmonizing visual text comprehension and generation,

    Z. Zhao, J. Tang, B. Wu, C. Lin, S. Wei, H. Liu, X. Tan, Z. Zhang, C. Huang, and Y . Xie, “Harmonizing visual text comprehension and generation,”arXiv preprint arXiv:2407.16364, 2024

    Pith/arXiv arXiv 2024

  45. [45]

    Pargo: Bridging vision-language with partial and global views,

    A.-L. Wang, B. Shan, W. Shi, K.-Y . Lin, X. Fei, G. Tang, L. Liao, J. Tang, C. Huang, and W.-S. Zheng, “Pargo: Bridging vision-language with partial and global views,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 7, 2025, pp. 7491–7499

    2025

  46. [46]

    Attentive eraser: Unleashing diffusion model’s object removal potential via self-attention redirection guidance,

    W. Sun, X.-M. Dong, B. Cui, and J. Tang, “Attentive eraser: Unleashing diffusion model’s object removal potential via self-attention redirection guidance,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 19, 2025, pp. 20 734–20 742

    2025

  47. [47]

    A bounding box is worth one token: Interleaving layout and text in a large language model for document understanding,

    J. Lu, H. Yu, Y . Wang, Y . Ye, J. Tang, Z. Yang, B. Wu, Q. Liu, H. Feng, H. Wanget al., “A bounding box is worth one token: Interleaving layout and text in a large language model for document understanding,”arXiv preprint arXiv:2407.01976, 2024

    Pith/arXiv arXiv 2024

  48. [48]

    Tabpedia: Towards comprehensive visual table understanding with concept synergy,

    W. Zhao, H. Feng, Q. Liu, J. Tang, B. Wu, L. Liao, S. Wei, Y . Ye, H. Liu, W. Zhouet al., “Tabpedia: Towards comprehensive visual table understanding with concept synergy,” Advances in Neural Information Processing Systems, vol. 37, pp. 7185–7212, 2025

    2025

  49. [49]

    Textsquare: Scaling up text-centric visual instruction tuning,

    J. Tang, C. Lin, Z. Zhao, S. Wei, B. Wu, Q. Liu, H. Feng, Y . Li, S. Wang, L. Liao et al., “Textsquare: Scaling up text-centric visual instruction tuning,”arXiv preprint arXiv:2404.12803, 2024

    Pith/arXiv arXiv 2024

  50. [50]

    Unidoc: A universal large multimodal model for simultaneous text detection, recognition, spotting and understanding,

    H. Feng, Z. Wang, J. Tang, J. Lu, W. Zhou, H. Li, and C. Huang, “Unidoc: A universal large multimodal model for simultaneous text detection, recognition, spotting and understanding,” arXiv preprint arXiv:2308.11592, 2023

    Pith/arXiv arXiv 2023

  51. [51]

    Docpedia: Unleashing the power of large multimodal model in the frequency domain for versatile document understanding,

    H. Feng, Q. Liu, H. Liu, J. Tang, W. Zhou, H. Li, and C. Huang, “Docpedia: Unleashing the power of large multimodal model in the frequency domain for versatile document understanding,”Science China Information Sciences, vol. 67, no. 12, pp. 1–14, 2024

    2024

  52. [52]

    Prolonged reasoning is not all you need: Certainty-based adaptive routing for efficient llm/mllm reasoning,

    J. Lu, H. Yu, S. Xu, S. Ran, G. Tang, S. Wang, B. Shan, T. Fu, H. Feng, J. Tanget al., “Prolonged reasoning is not all you need: Certainty-based adaptive routing for efficient llm/mllm reasoning,”arXiv preprint arXiv:2505.15154, 2025

    Pith/arXiv arXiv 2025

  53. [53]

    Advancing sequential numerical prediction in autoregressive models,

    X. Fei, J. Lu, Q. Sun, H. Feng, Y . Wang, W. Shi, A.-L. Wang, J. Tang, and C. Huang, “Advancing sequential numerical prediction in autoregressive models,”arXiv preprint arXiv:2505.13077, 2025

    Pith/arXiv arXiv 2025

  54. [54]

    Lightrag: Simple and fast retrieval-augmented generation,

    Z. Guo, L. Xia, Y . Yu, T. Ao, and C. Huang, “Lightrag: Simple and fast retrieval-augmented generation,” 2024. [Online]. Available: https://arxiv.org/abs/2410.05779

    Pith/arXiv arXiv 2024

  55. [55]

    Communitykg-rag: Leveraging community structures in knowledge graphs for advanced retrieval-augmented generation in fact-checking,

    R.-C. Chang and J. Zhang, “Communitykg-rag: Leveraging community structures in knowledge graphs for advanced retrieval-augmented generation in fact-checking,” 2024. [Online]. Available: https://arxiv.org/abs/2408.08535 25

    Pith/arXiv arXiv 2024

  56. [56]

    Afrikaans, inc.: the afrikaans culture industry after apartheid,

    A. S. Steyn, “Afrikaans, inc.: the afrikaans culture industry after apartheid,” Social Dynamics, vol. 42, pp. 481 – 503, 2016. [Online]. Available: https: //api.semanticscholar.org/CorpusID:152269054

    2016

  57. [57]

    Llms for low resource languages in multilingual, multimodal and dialectal settings,

    F. Alam, S. A. Chowdhury, S. Boughorbel, and M. Hasanain, “Llms for low resource languages in multilingual, multimodal and dialectal settings,” inConference of the European Chapter of the Association for Computational Linguistics, 2024. [Online]. Available: https://api.semanticscholar.org/CorpusID:268417133

    2024

  58. [58]

    Self-preference bias in llm-as-a-judge,

    K. Wataoka, T. Takahashi, and R. Ri, “Self-preference bias in llm-as-a-judge,”arXiv preprint arXiv:2410.21819, 2024

    Pith/arXiv arXiv 2024

  59. [59]

    Enhancing thyroid disease prediction using machine learning: A comparative study of ensemble models and class balancing techniques,

    J. Zhong and Y . Wang, “Enhancing thyroid disease prediction using machine learning: A comparative study of ensemble models and class balancing techniques,” 2025

    2025

  60. [60]

    Blind image quality assessment via vision-language correspondence: A multitask learning perspective,

    W. Zhang, G. Zhai, Y . Wei, X. Yang, and K. Ma, “Blind image quality assessment via vision-language correspondence: A multitask learning perspective,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 14 071– 14 081

    2023

  61. [61]

    You can even annotate text with voice: Transcription-only-supervised text spotting,

    J. Tang, S. Qiao, B. Cui, Y . Ma, S. Zhang, and D. Kanoulas, “You can even annotate text with voice: Transcription-only-supervised text spotting,” inProceedings of the 30th ACM International Conference on Multimedia, ser. MM ’22. New York, NY , USA: Association for Computing Machinery, 2022, p. 4154–4163. [Online]. Available: https://doi.org/10.1145/35031...

    arXiv 2022

  62. [62]

    Ocrbench v2: An improved benchmark for evaluating large multimodal models on visual text localization and reasoning,

    L. Fu, B. Yang, Z. Kuang, J. Song, Y . Li, L. Zhu, Q. Luo, X. Wang, H. Lu, M. Huanget al., “Ocrbench v2: An improved benchmark for evaluating large multimodal models on visual text localization and reasoning,”arXiv preprint arXiv:2501.00321, 2024

    Pith/arXiv arXiv 2024

  63. [63]

    Paddleocr: A versatile ocr toolkit with 80+ languages recognition,

    P. Authors, “Paddleocr: A versatile ocr toolkit with 80+ languages recognition,” https: //github.com/PaddlePaddle/PaddleOCR, 2023

    2023

  64. [64]

    Seed1. 5-vl technical report,

    D. Guo, F. Wu, F. Zhu, F. Leng, G. Shi, H. Chen, H. Fan, J. Wang, J. Jiang, J. Wanget al., “Seed1. 5-vl technical report,”arXiv preprint arXiv:2505.07062, 2025

    Pith/arXiv arXiv 2025

  65. [65]

    Vision as lora,

    H. Wang, Y . Ye, B. Li, Y . Nie, J. Lu, J. Tang, Y . Wang, and C. Huang, “Vision as lora,” arXiv preprint arXiv:2503.20680, 2025

    Pith/arXiv arXiv 2025

  66. [66]

    Cme-cad: Heterogeneous collaborative multi-expert reinforcement learning for cad code generation,

    K. Niu, H. Yu, Z. Chen, Z. Yao, W. Jia, X. Ge, J. Tang, B. Cui, B. Li, and X. Xue, “Cme-cad: Heterogeneous collaborative multi-expert reinforcement learning for cad code generation,”arXiv preprint arXiv:2512.23333, 2025

    arXiv 2025

  67. [67]

    Meml-grpo: Heterogeneous multi-expert mutual learning for rlvr advancement,

    W. Jia, J. Lu, H. Yu, S. Wang, G. Tang, A.-L. Wang, W. Yin, D. Yang, Y . Nie, B. Shan et al., “Meml-grpo: Heterogeneous multi-expert mutual learning for rlvr advancement,” arXiv preprint arXiv:2508.09670, 2025

    arXiv 2025

  68. [68]

    Mindev: Multi-modal integrated diffusion framework for video reconstruction from eeg signals,

    S. Huang, Y . Wang, H. Luo, H. Jing, C. Qin, and J. Tang, “Mindev: Multi-modal integrated diffusion framework for video reconstruction from eeg signals,” inProceedings of the 33rd ACM International Conference on Multimedia, 2025, pp. 3350–3359. 26

    2025

  69. [69]

    Resolving evidence spar- sity: Agentic context engineering for long-document understanding,

    K. Liu, Z. Chen, M. Li, J. Tang, D. Yang, and L. Zhang, “Resolving evidence spar- sity: Agentic context engineering for long-document understanding,”arXiv preprint arXiv:2511.22850, 2025

    arXiv 2025

  70. [70]

    Dolphin: Document image parsing via heterogeneous anchor prompting,

    H. Feng, S. Wei, X. Fei, W. Shi, Y . Han, L. Liao, J. Lu, B. Wu, Q. Liu, C. Linet al., “Dolphin: Document image parsing via heterogeneous anchor prompting,”arXiv preprint arXiv:2505.14059, 2025

    Pith/arXiv arXiv 2025

  71. [71]

    Wilddoc: How far are we from achieving comprehensive and robust document understanding in the wild?

    A.-L. Wang, J. Tang, L. Lei, H. Feng, Q. Liu, X. Fei, J. Lu, H. Wang, W. Liu, H. Liu et al., “Wilddoc: How far are we from achieving comprehensive and robust document understanding in the wild?”arXiv preprint arXiv:2505.11015, 2025

    Pith/arXiv arXiv 2025

  72. [72]

    Enhancing code llms with reinforcement learning in code generation,

    J. Wang, Z. Zhang, Y . He, Y . Song, T. Shi, Y . Li, H. Xu, K. Wu, G. Qian, Q. Chenet al., “Enhancing code llms with reinforcement learning in code generation,”arXiv preprint arXiv:2412.20367, 2024

    Pith/arXiv arXiv 2024

  73. [73]

    Bench- marking vision-language models on chinese ancient documents: From ocr to knowledge reasoning,

    H. Yu, Y . Wu, F. Shi, L. Liao, J. Lu, X. Ge, H. Wang, M. Zhuo, X. Wu, X. Feiet al., “Bench- marking vision-language models on chinese ancient documents: From ocr to knowledge reasoning,”arXiv preprint arXiv:2509.09731, 2025

    Pith/arXiv arXiv 2025

  74. [74]

    Jack and the beanstalk: Towards question answering in plant biology

    G. Jhajj and Y . Nomura, “Jack and the beanstalk: Towards question answering in plant biology.” [Online]. Available: https://api.semanticscholar.org/CorpusID:274567831

  75. [75]

    Long-context llms struggle with long in-context learning,

    T. Li, G. Zhang, Q. D. Do, X. Yue, and W. Chen, “Long-context llms struggle with long in-context learning,” 2024. [Online]. Available: https://arxiv.org/abs/2404.02060

    Pith/arXiv arXiv 2024

  76. [76]

    Rag based question-answering for contextual response prediction system,

    S. Veturi, S. Vaichal, R. L. Jagadheesh, N. I. Tripto, and N. Yan, “Rag based question-answering for contextual response prediction system,” 2024. [Online]. Available: https://arxiv.org/abs/2409.03708

    Pith/arXiv arXiv 2024

  77. [77]

    Fine-grained heartbeat waveform monitoring with rfid: A latent diffusion model,

    Y . Wang, T. Zhao, and X. Wang, “Fine-grained heartbeat waveform monitoring with rfid: A latent diffusion model,” inProceedings of the 3rd International Workshop on Human-Centered Sensing, Modeling, and Intelligent Systems, 2025, pp. 86–91. 27

    2025

  78. [2024]

    Available: https://arxiv.org/abs/2408.04187

    [Online]. Available: https://arxiv.org/abs/2408.04187

    Pith/arXiv arXiv