# ETD-OD: Object Detection for Parsing Electronic Theses and Dissertations > Research notes on ETD-OD, a 25K-page dataset with 24 element categories for layout analysis of electronic theses and dissertations, plus a PDF-to-XML parsing pipeline using object detection. - Authors: Hunter Heidenreich - Published: 2026-03-14 - Updated: 2026-03-15 - Paper: Published (WIESP 2022) - Project: ETD-OD (dataset + code + models) - Source: Bevaya Labs, --- ## TL;DR Ahuja et al. introduce ETD-OD, a 25K-page dataset of electronic theses and dissertations (ETDs) annotated with bounding boxes across 24 element categories, along with a complete PDF-to-XML parsing pipeline. YOLOv7 achieves 85.3 mAP@0.50, and pre-training on DocBank followed by fine-tuning on ETD-OD provides large gains for Faster R-CNN (+37.1 mAP@0.50). ## What kind of paper is this? **Dominant:** $\Psi\_{\text{Resource}}$ . The headline contribution is the ETD-OD dataset: 25K page images from 200 ETDs with ~100K bounding-box annotations across 24 document element categories. The dataset targets a document type (theses/dissertations) not covered by existing layout datasets. **Secondary:** $\Psi\_{\text{Method}}$ . The paper also presents an end-to-end PDF-to-XML parsing pipeline (object detection, text extraction, caption-to-figure matching, XML structuring) and benchmarks four detector configurations. ## What is the motivation? Electronic theses and dissertations contain substantial scholarly knowledge but are locked in unstructured PDFs, limiting their use in search, retrieval, summarization, and question-answering systems. Existing document layout datasets and parsing tools focus on shorter research papers (PubLayNet, DocBank, GROBID), which differ from ETDs in several ways: - ETDs are long-form documents (hundreds of pages) with distinct structural elements: committee pages, degree metadata, lists of contents, and back matter. - ETDs lack universal formatting; different institutions and disciplines impose different requirements and include domain-specific elements (algorithms in CS, equations in mathematics). - Layout differences (single vs. double column, wider margins, larger font sizes) mean models trained on research paper datasets do not transfer well. No existing dataset addressed these gaps at the time of publication. ## What is the novelty? ### ETD-OD Dataset The dataset contains 25,073 page images from 200 ETDs sampled across institutions, degree types, and academic domains. Pages are annotated with bounding boxes across 24 categories organized into five groups: | Group | Categories | | :--- | :--- | | **Metadata** (6) | Title, Author, Date, University, Committee, Degree | | **Abstract** (2) | Abstract Heading, Abstract Text | | **List of Contents** (2) | LOC Heading, LOC Text | | **Main Content** (12) | Chapter Title, Section, Paragraph, Figure, Figure Caption, Table, Table Caption, Equation, Equation Number, Algorithm, Footnote, Page Number | | **Bibliography** (2) | Reference Heading, Reference Text | Total annotations: 99,859 bounding boxes. Paragraphs (30,359) and page numbers (24,543) dominate; algorithms (96) are the rarest class. Annotation was performed by 6 undergraduate students using Roboflow, with each sample validated by 2 graduate students. ### PDF-to-XML Pipeline The parsing pipeline has three stages: 1. **Preprocessing:** Convert PDF pages to images using `pdf2image`. 2. **Element extraction:** Apply an object detection model (Faster R-CNN or YOLO) to produce bounding boxes with class labels. Post-processing rules correct common misclassifications (e.g., abstract heading vs. chapter heading via keyword matching and position constraints). 3. **XML structuring:** Image-based objects (figures, tables, algorithms, equations) are cropped and stored as files. Text-based objects are extracted via `pymupdf` (born-digital) or `pytesseract` (scanned). Captions are matched to figures/tables by Euclidean distance between bounding boxes. Output follows a hierarchical XML schema: front matter, body (chapters with sections), back matter (references). ## What experiments were performed? Four object detection configurations were trained and evaluated on an 80/20 train/validation split: | Model | mAP@0.50 | mAP@0.50:0.95 | | :--- | :--- | :--- | | Faster R-CNN (vanilla, ResNeXt-101) | 39.1 | 19.6 | | Faster R-CNN (DocBank pre-trained) | 76.2 | 44.0 | | YOLOv5 | 83.4 | 52.1 | | **YOLOv7** | **85.3** | **52.7** | Faster R-CNN models were trained for 60K iterations using Detectron2 (inference threshold 0.7). YOLO models were trained for 150 epochs. ### Per-Category Results (YOLOv7, AP@0.50) Strong categories (AP > 95): Paragraph (97.4), LOC Text (99.3), Reference Text (99.3), Figure (98.4), Footnote (98.9). Weak categories: Page Number (51.3), Equation Number (55.0), Algorithm (66.6), Date (68.3), Degree (68.3). The authors attribute low performance to limited training samples (degree, date, algorithm) and small object sizes (page numbers, equation numbers). ### Cross-Dataset Transfer A Faster R-CNN experiment on shared categories between DocBank and ETD-OD tested three training configurations: - **DocBank only:** Poor transfer to ETDs due to layout differences. - **ETD-OD only:** Reasonable but not optimal. - **DocBank + ETD-OD (pre-train + fine-tune):** Best results across all shared categories. This confirms that domain-specific fine-tuning on ETD data is necessary, and that scholarly document pre-training provides a useful initialization. ## What are the outcomes/conclusions? **Key findings:** - YOLOv7 achieves the best overall performance (85.3 mAP@0.50) on ETD-OD, outperforming Faster R-CNN substantially. - Pre-training on scholarly documents (DocBank) nearly doubles Faster R-CNN's mAP (39.1 to 76.2), confirming that domain overlap matters for document layout detection. - The proposed XML schema and parsing pipeline provide a complete end-to-end system for converting ETD PDFs to structured representations. - ETDs present challenges distinct from shorter research papers: metadata diversity, class imbalance (algorithms appear on < 0.4% of pages), and layout variability across institutions and disciplines. **Limitations and open questions:** - **Class imbalance is severe.** Algorithm (96 instances) and several metadata classes have very few annotations. The authors acknowledge this but do not apply mitigation strategies (oversampling, class-weighted loss). - **No formal inter-annotator agreement.** Although each sample was validated by a graduate student, no IAA metric (e.g., Krippendorff's Alpha or IoU agreement) is reported. - **Generalization is untested.** All data comes from US institutional repositories. Performance on non-English ETDs or ETDs from non-US institutions is unknown. - **The XML pipeline is not evaluated end-to-end.** The paper benchmarks object detection separately but does not measure the quality of the final XML output (e.g., text extraction accuracy, caption-to-figure matching precision). - **Small object detection remains a challenge.** Page numbers and equation numbers have AP below 56, and the paper does not explore specialized techniques for small objects (e.g., high-resolution inputs, feature pyramid tuning). - **Only standard COCO metrics are reported.** No error bars, multiple runs, or significance tests are provided. ## Reproducibility ### Models - **Faster R-CNN:** ResNeXt-101 backbone, implemented via Detectron2. - **Faster R-CNN (DocBank):** Pre-trained on DocBank, fine-tuned on ETD-OD. Uses the original DocBank model zoo weights and configurations. - **YOLOv5:** Open-source implementation; the paper cites Jocher et al. 2022 (ultralytics/yolov5 v6.2) but does not pin a specific commit or version in the repo. - **YOLOv7:** Open-source implementation; specific version not stated. - **Weights availability:** The paper claims pre-trained models are available at the GitHub repository. However, as of March 2026, the repository (https://github.com/Opening-ETDs/ETD-OD) is not publicly accessible. The paper uses the URL `Opening-ETDS` (capital S) while the note uses `Opening-ETDs`; neither variant resolves to a public repository. Model weights are effectively unavailable. ### Algorithms - Faster R-CNN: 60K iterations, inference threshold 0.7. No learning rate, optimizer, batch size, or augmentation details are reported. - YOLO models: 150 epochs. No other training hyperparameters (learning rate, optimizer, augmentation, image resolution) are reported. - Post-processing: Rule-based corrections for abstract headings (keyword matching + position constraint: chapter heading in first 10 pages matching "abstract" keyword) and element-to-caption matching (Euclidean distance between bounding boxes, with a y-coordinate constraint for equation numbers). ### Data - **Source:** 200 ETDs from publicly accessible institutional repositories. Sampling stratified by degree, domain, and institution, though specific repositories are not named. - **Annotation:** Roboflow platform. 6 undergraduate annotators + 2 graduate reviewers. No inter-annotator agreement metrics reported. - **Split:** 80/20 train/validation. No separate test set. - **Availability:** The paper states that dataset, code, and models are available at a GitHub repository. As of March 2026, this repository is not publicly accessible (neither the `Opening-ETDS` nor `Opening-ETDs` URL variant resolves). No dataset files, annotation files, or code have been released publicly. ### Evaluation - COCO-style mAP@0.50 and mAP@0.50:0.95. - Evaluation on validation split only; no held-out test set. - No cross-dataset evaluation (e.g., testing on PubLayNet or DocLayNet). - No statistical rigor: single runs, no error bars, no significance tests. - Evaluation scripts are not available (no code in the repository). ### Hardware - Not reported. No mention of GPU type, training time, memory requirements, or inference throughput. ## BibTeX ```bibtex @inproceedings{ahuja-etal-2022-parsing, title = {Parsing Electronic Theses and Dissertations Using Object Detection}, author = {Ahuja, Aman and Devera, Alan and Fox, Edward Alan}, booktitle = {Proceedings of the First Workshop on Information Extraction from Scientific Publications}, month = nov, year = {2022}, address = {Online}, publisher = {Association for Computational Linguistics}, doi = {10.18653/v1/2022.wiesp-1.14}, pages = {121--130}, } ```