acceptodds
Withdrawn from ICLR 2027

DocLayout-YOLO: Axis-Attentive Document Layout Detector with Mesh-Guided Synthetic Pretraining

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Accurate and efficient document layout analysis is critical for modern document parsing systems. However, although the YOLO framework represents a leading paradigm for the real-time detector, its adaptation to heterogeneous documents remains underexplored. To address this gap, we propose DocLayout-YOLO, a document-specialized real-time detector that integrates diverse synthetic document pretraining and page-axis context modeling. From the data perspective, we introduce MeshLayout-1M, a synthetic pretraining corpus generated by the proposed Mesh-candidate BestFit algorithm, which provides large-scale and diverse document layouts for pretraining and improves adaptation to heterogeneous document scenarios. From the model perspective, we propose Page-Axis Layout Attention (PALA), which captures long-range dependencies along the horizontal and vertical page axes while remaining computationally efficient. Extensive experiments demonstrate that DocLayout-YOLO achieves state-of-the-art performance on four heterogeneous document benchmarks, obtaining 72.4%/75.5%/96.0%/73.1% mAP on DLA, MDoc, HJDataset, and PRIMA-LAD, respectively. Compared with existing real-time detectors, DocLayout-YOLO establishes a superior accuracy-efficiency trade-off in document layout analysis.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Object Detection, Document Layout Analysis

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_K38vOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This study proposes a novel real-time Document Layout Analysis (DLA) model. The proposed model is based on the YOLOv12 architecture. One of the core contributions is replacing the base architecture's Area Attention block with Page-Axis Layout Attention (PALA). PALA performs full attention along horizontal and vertical axes rather than the block attention used in Area Attention. PALA leverages the inductive bias that document layouts exhibit strong horizontal and vertical structure. Another core contribution is a synthetic document generation method called Mesh-candidate BestFit (MBF). The study shows that synthetic document data generated by MBF improve model performance. The study conducted extensive experiments on representative DLA benchmarks and showed that the model achieved superior performance to existing real-time detectors and specialized DLA methods.

Critical strengths
  1. The study conducted extensive comparisons across various benchmarks and alternative methods.
  2. The study analyzes performance from multiple perspectives including accuracy and speed.
  3. The study conducted sufficient ablation studies on the proposed core components.
  4. Strong performance.
Critical weaknesses

Major Weaknesses

  1. MBF is one of the core contributions of this study. 1M synthetic pages generated by MBF contribute to improving model performance. However, I could not find any statement about releasing the MBF source code in the current manuscript. Although methodological details are described in Algorithm 1, it is strongly recommended to release the generation code for reproducibility and follow-up studies.
  2. YOLO-DLA [1] is a document-specific YOLO-based method and uses axial attention which seems closely related to PALA. It is worth discussing as a related method. An additional comparison would also be valuable if feasible.
  3. The novelty of PALA and its relationship to existing axis-aware methods should be positioned more clearly in the main text. Approaches exploiting horizontal and vertical structure already exist. The Introduction does not sufficiently explain how PALA differs from existing axis-aware methods. PALA's distinguishing features, such as parallel H/V attention, independent normalization, and full-axis interaction, should be described clearly in the Introduction or Related Work.
  4. DocLayNet is a widely used DLA benchmark. Although comparison results are reported in the Appendix C.1, they are limited. I recommend adding comparisons with recent DLA methods to Table 3. If such comparisons are not feasible, the reason should be explained.
  5. The effectiveness of synthetic data generated by MBF is evaluated only on M6Doc\text{M}^6\text{Doc} in Table 5. It would be valuable to include another benchmark, such as PRIMA-LAD. This would provide a more complete evaluation of its generalizability.

Minor Weaknesses and Suggestions

  1. This study provides sufficient aggregate ablation studies on PALA's effectiveness. In addition, a sample-wise qualitative analysis showing how the horizontal and vertical branches capture structural relationships in real document samples would be beneficial for understanding the behavior of the proposed method.
  2. DocMY/Mamba-YOLO [2] overlaps with several benchmarks used in this study. I recommend adding quantitative comparison results to Table 3.
  3. Please correct the following typos:
  • L824: deault -> default
  • L1059: EVALUATION EVALUATION -> EVALUATION
  1. I recommend reporting the generation speed of synthetic document generation methods, such as the number of pages generated per second.

[1] Qi, Haoyan, Xinyang Meng, and Zhijuan Du. "Yolo-dla: A yolo-based unified framework for multi-scale document layout analysis." Expert Systems with Applications (2025): 129981.

[2] Ma, Wenkang, et al. "Mamba-YOLO: Multi-level adaptive rectangular convolution for document layout analysis." Pattern Recognition 170 (2026): 112031.

Questions for authors
  1. Table 1 states that PRIMA-LAD contains 478 pages, 7453 elements, and 5 element types. However, the footnote in Appendix B.1 states that "11-class annotation configuration adopted in our experiments, containing 8,942 valid regions over 478 pages". Could the authors clarify this discrepancy?
Reviewer_V4YnOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DocLayout-YOLO, a document layout detector built on YOLOv12, combines MeshLayout-1M, a synthetic pretraining dataset generated by Mesh-candidate BestFit, with Page-Axis Layout Attention (PALA), which aggregates context along the horizontal and vertical page axes. The model achieves reported mAP₅₀:₉₅ scores of 72.4, 75.5, 96.0, and 73.1 on D⁴LA, M⁶Doc, HJDataset, and PRIMA-LAD, respectively.

Critical strengths
  1. On M⁶Doc, DocLayout-YOLO-X improves mAP from 74.0 to 75.5 while reducing latency from 8.5 ms to 6.9 ms relative to YOLOv12-X. Results across multiple model scales and seven benchmarks support its practical applicability. The five-scale ablation study separates the contributions of synthetic pretraining and PALA.
Critical weaknesses
  1. The proposed models receive substantial synthetic pretraining, whereas several YOLO baselines are trained from scratch. In Table 4, the authors address this difference for YOLOv12 but not for other competitive detectors.

  2. CSWin already uses parallel horizontal and vertical attention with separate normalization, so PALA’s contribution should be distinguished through its channel allocation and fusion design.

Questions for authors
  1. Reported FLOPs exclude attention aggregation, which is precisely the computation modified by PALA. It would be better also report this part.
Reviewer_Uyn7OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents DocLayout-YOLO, a YOLOv12-based detector for document layout analysis. It introduces two main improvements:

(1) a layout synthesis method that generates diverse document pages for pretraining to improve downstream detection performance;

(2) a new attention method called Page-Axis Layout Attention (PALA), which captures long-range horizontal and vertical dependencies through parallel attention branches with separate normalization.

Experiments across multiple document benchmarks demonstrate improved detection accuracy and a favorable accuracy–efficiency trade-off.

Critical strengths
  1. Relevant motivation: Document layout analysis requires finer-grained distinctions than identifying visually distinct objects in generic detection. Beyond figures, equations, and tables, it involves distinguishing closely related categories such as figure captions and table captions. The paper addresses this need through document-specific data synthesis and attention design.

  2. Clear presentation and supporting experiments: The paper clearly motivates its two improvements and supports both with internal ablation studies examining their design choices. Experiments across multiple document benchmarks further demonstrate competitive detection accuracy and a favorable accuracy–efficiency trade-off.

Critical weaknesses
  1. Limited comparison with closely related layout detectors: Although the paper includes internal ablations and comparisons with general-purpose detectors and specialized DLA methods, it omits relevant efficiency-oriented models such as PP-DocLayout. Direct accuracy–latency comparisons under consistent evaluation settings would better substantiate its claimed advantages.

  2. Positioning relative to grounding-capable VLMs: A comparison with representative grounding-capable VLMs would make the evaluation more complete and help clarify the benefits of a specialized detector in terms of fine-grained localization, accuracy, and computational cost.

  3. Further analysis of the synthesis strategy: Although several synthesis baselines are included, comparisons with matched data volumes and training budgets would better isolate the benefits of the proposed generation method. Its applicability to photographed or degraded documents also remains insufficiently established, particularly under perspective distortion, uneven lighting, and blur. Targeted evaluations would help clarify these limitations.

  4. Reproducibility and release details: The paper should clarify whether MeshLayout-1M and the complete synthesis pipeline will be publicly released, including preprocessing scripts and generation configurations. Their availability would support reproducibility and enable the community to extend the approach to other document domains.

Questions for authors

Please refer to the weaknesses above. I would appreciate further clarification on the suggested comparisons and the limitations of the synthesis strategy. Please also clarify the release plans for the synthetic dataset and generation tools.

Reviewer_gaHoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies efficient document layout detection through synthetic document pretraining and page-axis context modeling. The authors introduce MeshLayout-1M, generated using Mesh-candidate BestFit, and Page-Axis Layout Attention (PALA), which replaces YOLOv12’s Area Attention with parallel horizontal and vertical attention branches using separate normalization. The resulting DocLayout-YOLO is evaluated on four main benchmarks across five model scales, with additional experiments on other datasets, robustness, pretraining scale, and component effectiveness.

Critical strengths

Practical relevance supported by broad empirical evaluation. Efficient document layout detection is valuable for document processing systems. The paper evaluates four main benchmarks and five model scales, reporting improvements in both accuracy and measured inference latency. For example, on M6Doc, the X-scale model improves mAP from 74.0 to 75.5 relative to YOLOv12-X while reducing TensorRT FP16 latency from 8.5 ms to 6.9 ms. These gains make the overall system potentially useful in practice.

Substantive controlled analysis of the proposed components. Table 4 includes a complete factorial ablation of pretraining and PALA, allowing their contributions to be distinguished within YOLOv12. Table 7 and Appendix D further compare attention mechanisms and execution/normalization strategies. The acknowledgment that PALA is not theoretically more efficient than every possible Area Attention configuration also provides a useful qualification of its efficiency advantage.

Critical weaknesses

The cross-method comparison in Table 3 does not consistently use comparable evaluation metrics. Appendix A states that results for specialized DLA methods are taken from their original publications. A concrete example is DocSAM: the reported D4LA/M6Doc/HJDataset scores of 55.5/50.0/90.5 correspond to mask mAP in Table 4 of the DocSAM paper, which explicitly distinguishes this metric from bounding-box mAP. Moreover, Table 5 of that paper reports 66.3 bounding-box mAP on M6Doc, whereas this submission uses 50.0 in its comparison. Thus, the current table combines different evaluation targets and does not use the stronger published detection result for this baseline. This requires correction before the cross-method claims can be assessed reliably. The proposed method may still outperform the corrected baseline, but the current comparison does not establish that conclusion on a consistent basis. See DocSAM, CVPR 2025, Section 4.1 and Tables 4–5. Comparability on PRIMA-LAD also requires clarification. The experiments use an 11-class configuration, while the dataset overview lists five classes. It is unclear whether the quoted baseline results use the same label mapping, data split, and evaluation protocol.

The computational-efficiency comparison omits the central operation changed by the method. Page 6 explicitly states that the FLOPs profiling excludes attention aggregation, resulting in identical reported FLOPs for PALA and YOLOv12. Since attention aggregation is precisely the operation being modified, the FLOPs–accuracy plot does not constitute a complete computational comparison. The latency measurements under a common H200/TensorRT environment remain useful evidence, but they exclude data transfer, preprocessing, decoding, and NMS. The paper should report complete computational costs and explicitly scope the measured speed advantage to engine execution. This matters because the accuracy–efficiency trade-off is a central contribution.

Questions for authors

Please identify whether each result in Table 3 uses bounding-box AP or mask AP and provide a comparison using consistent metrics. In particular, why were DocSAM’s Table 4 mask scores used instead of its published M6Doc detection result in Table 5? Which principal conclusions remain supported after correction?

Do the quoted PRIMA-LAD baselines use the same 11-class annotations, training/test split, and evaluation protocol as this work? If not, please provide comparable results or appropriately qualify the associated claims.

What are the complete FLOPs of YOLOv12 and PALA when attention aggregation is included? Please update the relevant comparisons and distinguish engine-execution latency from end-to-end page-processing latency.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.