Datasets/Softwares

Circuit diagram:

Floor plan:

  • BRIDGE: Building Plan Repository for Image Description Generation, and Evaluation (paper)
  • CVC-FP (Database for structural floor plan analysis)
  • FPLAN-POLY dataset of vectorized graphic documents (floorplans)
  • SESYD: Systems Evaluation SYnthetic Documents11 types of synthetic documents (paper)
  • R-FP-500: Floor plan from Rakuten Real Estate and pixel-wise wall label (by Rakuten Institute of Technology)

Maps/cadastral:

  • Map Border Dataset: dataset for Detection and Segmentation tasks in Historical Cadastral Maps

Music Scores:

Comic book:

  • AI4VA: AI for Visual Arts dataset comprising comic-style imagery sourced from two mid-twentieth-century Franco-Belgian comics series, Placid et Muzo and Yves le loup for VLM benchmarking (arXiv, 2024).
  • BCBID: Bangla Comic Book Image Dataset contains a total of 3327 images of different kinds of ‘Bengali Comic Books’ from a diverse set of renowned authors (published at ICDAR 2019).
  • C3B: Comics Cross-Cultural Benchmark, a multicultural, multitask and multilingual cultural awareness capabilities benchmark. Comprises over 2000 images and over 18000 QA pairs (arXiv, 2025).
  • CDVSR: Comics Dataset for Visual Sentiment Recognition, 10,281​ images of comic and manga.
  • ChrOMIC: Chronological Reasoning in Multi-panel Comics is the first benchmark designed to evaluate vision-language models (VLMs) on their ability to understand panel ordering and narrative reasoning in comics (EACL 2026).
  • ComSet: 54K strips, harvested from 13 popular comics available online.
  • COMICORDA: A Novel Dataset for Dialogue Act Recognition in Comics, an extension of the EmoRecCom dataset.
  • COMICS: 1.2 million panels paired with automatic textbox transcriptions from Golden Age collection of the Digital Comics Museum (published at CVPR 2017). New OCRed text COMICS Text+ (2024).
  • Comics Datasets Framework: Mix of Comics datasets for detection benchmarking (ICDAR 2024)
  • ComicsPAP: understanding comic strips by picking the correct panel (arXiv 2025)
  • ComicScene154: focuses on scene-level narrative arcs annotation. It comprises four public-domain comic magazines containing 34 distinct stories that span a total of 154 pages from Golden Age public-domain American comics (arXiv, link).
  • COO: COmic Onomatopoeia dataset for recognizing arbitrary or truncated texts. Based on Manga109 images, it consists of 61,465 polygons and 2,261 links between truncated texts (ECCV 2022)
  • ComicVQA: a subset of the COMICS dataset supplemented with self-sourced 6 panel pages all associated with a 100-word description generated with GPT-4o-mini for each panel (ACL 2026)
  • DCM772: 772 annotated images from 27 Golden Age collection of the Digital Comics Museum. It includes ground-truth bounding boxes of all panels, all characters (body + faces), small or big, human-like or animal-like (published at MDPI Journal Imaging 2018).
  • eBDtheque: a representative database of comics of 100 pages including manual annotations of 850 panels and 1092 balloons paired with 1620 comic characters and 4693 text lines. (published at ICDAR 2013).
  • EmoComics35: a genre-diverse dataset consisting of 35 comic albums where utterances are annotated with character identity and fine-grained multi-class emotion labels (ICPR2026).
  • EmoRecCom: ICDAR2021 Competition Multimodal Emotion Recognition on Comics scenes (codalab) (ICDAR 2021).
  • FGC 2019: ICDAR 2019 Competition on Fine-Grained Classification of Comic Characters
  • GNC: the Graphic Narrative Corpus currently contains textual metadata of about 219 titles written in English. Corresponding image are not provided due to copyright issue (ICDAR 2017).
  • iCartoonFace: a large-scale challenging dataset established for cartoon face recognition. 389,678 images of 5,013 cartoon persons collected from 1,302 cartoon albums (published at ICM 2020).
  • IMCDB: Indian Mythological Comic Dataset – digitized Indian comic storybook in the English language (ICDAR 2021).
  • KABOOM ONOMATOPEA: Comic Onomatopoeia Dataset for Extracting Arbitrary or Truncated Texts
  • Large-scale Dataset for Robust Complex Anime Scene Text Detection: containing 735K images and 4.2M annotated text block position (arXiv)
  • Manga 109: 109 manga volumes from “Manga Library Z” drawn by professional manga artists in Japan (published in Multimedia Tools and Applications Journal 2017).
  • MangaSeg: 700,000 segmentation annotations for 10,130 double-sided manga pages from Manga 109 dataset. Panels, characters, faces, speech balloons, texts, and links between characters and balloon annotated at the instance-level (CVPR 2025).
  • MangaUB: a Manga Understanding Benchmark for Large Multimodal Models (IEEE MM25)
  • Mangazero: a large-scale, multi-role, multi-state comic image dataset containing 43,264 pages of comics and 427,147 annotated panels, which supports the visualization of various character interactions and actions in consecutive frames and is suitable for multi-character and multi-state comic generation tasks (CVPR 2025).
  • PopManga: includes 57,318 images from 100 of the most popular English manga. However, the full dataset is not publicly available; only a small test set of 1,925 images, collected from Manga Plus by Shueisha 2, has been released. Contains detection annotations and text-character and character-character links for dialog transcription (CVPR 2024).
  • OpenAI Comic Strips: 500 six-panel comic strips (3,000 images) generated with OpenAI’s gpt-image-1 (citation coming).
  • OpenMantra: evaluation dataset of 5 manga titles (JA/EN/ZH text+images) for machine translation, presented in AAAI 2021.
  • Re:Verse: a comprehensive benchmark designed to evaluate VLMs’ ability to understand long-form manga narratives (ICCV 2025).
  • Sequencity612: comic character annotation for all characters, small or big, speaking or not and in the background on 612 recent comic book pages (ICDAR 2017)
  • SSGCI 2016 ICPR 2016 Competition on Subgraph Spotting in Graph representation of Comic Book Images
  • STRIPCIPHER: dataset with 600 examples for prediction, 680 for comprehension and 890 for reordering (EMNLP 2025).
  • VLRC: Visual Language Research Corpus made up of ~36,000 coded panels from 300+ comics from Europe, Asia, and the United States, across time periods (1940-present), and various genres.
  • YManga: 1,015 high-quality yonkoma-type manga strips (EMNLP 2024).