# SciPostLayout: A Layout Dataset for Scientific Posters > SciPostLayout introduces the first publicly licensed dataset for layout analysis and generation of scientific conference posters, covering 7,855 annotated poster images across nine element categories. - Authors: Bevaya AI Team - Published: 2024-07-01 - Updated: 2026-03-12 - Dataset: SciPostLayout Dataset (HuggingFace) - Code: SciPostLayout Code (GitHub) - Paper: Preprint (arXiv) - Source: Bevaya Labs, --- ## TL;DR SciPostLayout is a dataset of 7,855 manually annotated scientific conference posters collected from F1000Research, all under a CC-BY license. It is the first publicly available dataset designed for layout analysis and layout generation in the scientific poster domain, which the authors find more challenging than existing scientific paper datasets like PubLayNet. ## What kind of paper is this? **Dominant:** $\Psi\_{\text{Resource}}$. The primary contribution is the SciPostLayout dataset itself: a curated, manually annotated collection of scientific poster images with bounding-box labels across nine element categories, released under an open license. The paper frames its experimental results as baseline benchmarks that accompany the dataset release. **Secondary:** $\Psi\_{\text{Evaluation}}$. The paper evaluates existing layout analysis and layout generation models on the new dataset to characterize the difficulty of the poster domain relative to prior scientific paper datasets. This benchmarking work is significant but subordinate to the dataset contribution. ## What is the motivation? Scientific conference posters efficiently communicate research findings in a graphical, space-constrained format. Creating a well-designed poster from a paper is labor-intensive work, and automating it is an appealing goal. Yet research on scientific poster generation remained sparse at the time of this paper, largely because no publicly available, properly licensed dataset existed for the task. Prior work on layout analysis (such as PubLayNet) focused on scientific papers in document form, not posters. Poster layouts differ structurally: figures and tables can appear at arbitrary positions, typography varies widely, and the visual density and arrangement patterns are quite different from journal article pages. Two prior datasets for poster generation existed, but one was not publicly available and the other had unclear licensing, making them unsuitable as shared benchmarks. This gap motivated the construction of a community resource that could support both layout analysis (detecting element regions in a poster image) and layout generation (producing plausible element arrangements), as well as the more ambitious paper-to-poster generation task. ## What is the novelty? The central contribution is the SciPostLayout dataset. Several aspects distinguish it from prior work: **Scope and domain.** All 7,855 items are scientific conference-style posters, not paper pages. This is an important distinction: poster layouts are substantially more variable than paper layouts because fonts, figure sizes, and column arrangements are not constrained by journal templates. **Annotation depth.** The authors expanded PubLayNet's five-category scheme to nine categories to capture the finer-grained structure of posters: | Category | Contents | |---|---| | Title | Paper title | | Author Info | Author names and affiliations | | Section | Top-level section headings | | Text | Body paragraphs | | List | Bullet points, numbered items, reference blocks | | Table | Table body | | Figure | Figure body (including sub-figure panels as a single bounding box) | | Caption | Captions for tables and figures | | Unknown | Large-area logos, advertising | **Paired paper-poster data.** The dataset includes 100 paper-poster pairs (50 dev, 50 test), all under CC-BY, enabling experiments on generating poster layouts directly from source papers. No prior publicly licensed dataset supported this. **Licensing.** All posters and paired papers are under CC-BY (versions 2.0, 3.0, or 4.0), making the dataset suitable for commercial research use. This was a deliberate design choice: the authors collected only from F1000Research and excluded any posters under noncommercial or non-distributable licenses. ## What experiments were performed? The paper benchmarks three tasks using existing models. ### Layout Analysis Two pre-trained document layout detection models were fine-tuned on SciPostLayout's training split and evaluated on the test split: - **LayoutLMv3** (with Cascade R-CNN detector) - **DiT** (with Cascade R-CNN detector) Both were initialized from base-size checkpoints and fine-tuned until dev-set performance peaked. The evaluation metric is mean average precision at IoU thresholds from 0.50 to 0.95, i.e., $\text{mAP}@\text{IoU}[0.50{:}0.95]$. The Unknown category was excluded due to insufficient instances. LayoutLMv3 achieved an overall mAP of 66.93; DiT reached 63.77. Both models performed well on Title and Author Info (which are structurally regular and appear at predictable locations on posters) but showed meaningful performance drops compared to their reported PubLayNet results, where mAP values exceed 90. The authors attribute this to the greater diversity of element positions and typography in posters. ### Layout Generation Three layout generation models were evaluated across five conditional generation settings: - **LayoutDM** (discrete diffusion, trained from scratch on SciPostLayout) - **LayoutFormer++** (transformer-based, trained from scratch) - **LayoutPrompter** (GPT-4-based, few-shot via API; no training needed) The generation settings vary by what information is given as a condition: - **Gen-T**: element type counts only - **Gen-TS**: element type counts and sizes - **Gen-R**: element types and pairwise relationships - **Completion**: a partial layout - **Refinement**: a noisy or suboptimal layout Evaluation metrics are: - **mIoU** (maximum IoU, $\uparrow$): the highest IoU between a generated layout and any real layout in the test set - **Alignment** ($\downarrow$): how well elements within a layout are aligned with each other - **Overlap** ($\downarrow$): the overlapping area between pairs of elements in a layout - **FID** (Frechet Inception Distance, $\downarrow$): distributional similarity between generated and real layouts in an embedding space FID and mIoU both measure similarity to the real layout distribution, but from different angles: mIoU is computed from geometric intersection between individual layouts, while FID is an embedding-based distributional metric. Their rankings across models are not always consistent. Results indicate that mIoU was low for all models and less than half the values reported on PubLayNet, confirming that poster layout generation is harder. All models produced well-aligned layouts (low Alignment scores). LayoutPrompter was the most effective at minimizing element overlap; LayoutDM produced layout distributions most similar to the real distribution (lowest FID) but with more overlap. In the Refinement setting, LayoutPrompter substantially outperformed the other models. ### Paper-to-Layout To explore automated poster generation from source papers, the authors implemented two settings using GPT-4: - **Gen-T**: GPT-4 extracts element type count constraints from the paper text (via Nougat for PDF parsing), then a layout generation model generates a layout as in the Gen-T generation experiment. - **Gen-P**: GPT-4 first summarizes the paper in under 1,000 words, then LayoutPrompter generates a layout from that summary using the 50 dev paper-poster pairs as few-shot examples. For Gen-T, constraint extraction accuracy was measured with mean absolute error (MAE) between predicted and actual element counts: $$\text{MAE} = \frac{1}{C} \sum\_{c=1}^{C} \left| n\_c^{\text{pred}} - n\_c^{\text{real}} \right|$$ where $C$ is the number of element categories and $n\_c$ is the element count for category $c$. GPT-4 achieved an overall MAE of 1.98, with the largest errors on Text (4.9) and Figure (3.66) categories, which are the most numerous and most variable. Layout generation quality degraded slightly compared to using ground-truth constraints, as expected. In the Gen-P setting (tested only with LayoutPrompter), the model produced well-aligned layouts with very low overlap (0.022), suggesting that LLMs have meaningful potential for this task even though they currently cannot reproduce the exact structure of real poster layouts. ## What are the outcomes/conclusions? The authors report that layout analysis on scientific posters is harder than on scientific paper images, with both LayoutLMv3 and DiT showing notable performance drops versus their PubLayNet benchmarks. Layout generation is similarly more challenging: mIoU values for all tested models are less than half of what is observed on PubLayNet. These results suggest that poster layout is a distinct and underexplored problem that warrants dedicated modeling effort. The paper-to-layout experiments are presented as a proof of concept. GPT-4 can extract plausible element constraints from paper text, and LayoutPrompter can then generate reasonably structured layouts. The generated layouts are not yet close to real poster layouts by quantitative measures, but the results indicate the task is tractable and worth further investment. A limitation that the authors acknowledge implicitly is domain skew: because posters were sourced exclusively from F1000Research, most of the collected posters are in the biomedical field. This may limit how well models trained on SciPostLayout transfer to posters from other scientific communities (e.g., physics, computer science conferences where PDF poster formats differ). ## Mapping to Matter vs. Meaning Framework How does SciPostLayout's 9-class poster taxonomy map to our [Matter vs. Meaning framework](../matter-vs-meaning/)? | SciPostLayout Class | Visual Primitive | Logical Role | Notes | | :--- | :--- | :--- | :--- | | **Title** | `Text` | `Title` | Paper title on the poster. | | **Author Info** | `Text` | `Author` / `Affiliation` | Author names and institutional affiliations, combined in a single class. | | **Section** | `Text` | `SectionHeader` | Top-level section headings on the poster. | | **Text** | `Text` | `Body` | Body paragraphs. | | **List** | `Text` | `ListItem` | Bullet points, numbered items, and reference blocks. Annotated as a container (not individual items). | | **Table** | `Table` | *(primitive only)* | Table body. | | **Figure** | `Image` | `Figure` / `Chart` / `Diagram` | Figure body, including sub-figure panels as a single bounding box. No sub-classification between photographs, charts, and diagrams. | | **Caption** | `Text` | `Caption` | Captions for tables and figures. | | **Unknown** | `Image` | `Logo` / `Advertisement` | Large-area logos and advertising elements. Excluded from evaluation due to insufficient instances. | **Coverage gaps:** SciPostLayout does not distinguish `Formula`, `PageHeader`, `PageFooter`, `Footnote`, `BibEntry`, or `Code`. Reference blocks are merged into the `List` class. The `Author Info` class combines author names and affiliations into a single label, unlike datasets that separate `Author` and `Affiliation`. The `Unknown` class is a catch-all for non-standard elements. ## Reproducibility ### Models No new models are released with this paper. The experiments use LayoutLMv3 (base, 133M parameters), DiT (base, 86M parameters), LayoutDM, LayoutFormer++, and LayoutPrompter (which calls the GPT-4 API, specifically gpt-4-1106-preview). Weights for LayoutLMv3 and DiT are publicly available from Microsoft; the layout generation models are available from their respective repositories. ### Algorithms LayoutLMv3 and DiT were fine-tuned from their base-size pretrained checkpoints using Cascade R-CNN as the object detection head. The checkpoint with the highest dev-set performance was selected for test evaluation. Specific optimizer settings, learning rates, and training durations are not reported in the paper. LayoutDM and LayoutFormer++ were trained from randomly initialized weights on SciPostLayout's training split. Detailed hyperparameters are not provided; readers would need to consult the original papers for those models. For paper-to-layout, Nougat was used to extract text from PDF files, and GPT-4 was prompted to extract element type constraints or generate a paper summary. The prompts used are reproduced in Appendix B of the paper. ### Data The dataset was constructed as follows: - Posters in PDF format were downloaded from F1000Research. - Only CC-BY licensed posters were retained (7,943 after license filtering). - PDFs were converted to PNG at DPI=100. - Posters with file sizes below 200KB were excluded as low-content (primarily text-only), leaving 7,855 posters. - Professional data annotators manually annotated bounding boxes and categories for all 7,855 posters. **Split:** 6,855 train / 500 dev / 500 test (from Table 2 in the paper; train annotation statistics sum to 168,313 total annotations, dev to 12,141, test to 12,243). **Paired papers:** 100 paper-poster pairs (50 dev, 50 test) were manually identified by searching for papers corresponding to posters in the dataset. There is no automated linking on F1000Research, so this was done by hand. **Domain skew:** The authors note through title-word analysis that most posters are from the biomedical field, which reflects F1000Research's coverage. The dataset is publicly available on HuggingFace at `omron-sinicx/scipostlayout_v2` under CC-BY. Note that older posters may carry CC-BY 2.0 or CC-BY 3.0 rather than 4.0. ### Evaluation **Layout analysis** is evaluated with $\text{mAP}@\text{IoU}[0.50{:}0.95]$ on bounding boxes. The Unknown category is excluded from evaluation. This follows standard COCO-style object detection evaluation. **Layout generation** uses four metrics: mIoU (maximum IoU between a generated layout and any real layout in the test set), Alignment, Overlap, and FID. These metrics are drawn from the layout generation literature and are not novel to this paper. **Paper-to-layout constraint extraction** is evaluated with MAE between predicted and actual element type counts, averaged over three GPT-4 inference runs. The paper compares against results reported in prior work on PubLayNet to contextualize difficulty, but direct cross-dataset comparison involves confounds (different split sizes, annotation granularity, domain). ### Hardware No hardware details are provided for the layout analysis or generation training runs. LayoutPrompter and the paper-to-layout experiments use GPT-4 via API, so inference cost depends on token usage rather than local hardware. ## BibTeX ```bibtex @inproceedings{tanaka2024scipostlayout, title = {SciPostLayout: A Dataset for Layout Analysis and Layout Generation of Scientific Posters}, author = {Tanaka, Shohei and Wang, Hao and Ushiku, Yoshitaka}, booktitle = {British Machine Vision Conference (BMVC)}, year = {2024} } ```