jaddai/openart-animals
收藏资源简介:
OpenArt — Animals & the Natural World(OpenArt动物与自然世界)是OpenArt系列开放公共领域艺术数据集的动物主题子集,包含31,511件作品(15,920幅绘画/插图、15,407件摄影对象、184件未分类),每件作品都配有结构化VLM描述以及媒介、归属和铭文元数据。该数据集涵盖了各种媒介中的动物和自然世界主题,包括鸟类、野兽、鱼类和昆虫的绘画和绘图,以及雕刻、铸造、釉面和编织成雕塑、陶瓷、金属制品和版画的作品。它是OpenArt中最大的主题子集。数据集来源自大都会艺术博物馆、荷兰国家博物馆、克利夫兰艺术博物馆、芝加哥艺术学院、史密森尼学会、Europeana、维基媒体等主要博物馆和图书馆的开放访问API。与仅包含绘画的OpenBrush语料库不同,每个OpenArt主题子集都是混合媒介的:包含二维艺术(绘画、版画、绘图)和三维对象的照片(雕塑、陶瓷、金属制品、纺织品)。通过brand列(openbrush表示二维艺术,openartifacts表示摄影对象,ambiguous表示未明确)可以精确筛选所需内容。关键特征包括:所有31,511张图像均为公共领域且通过rights_status: safe检查;每张图像具有9个语义部分的结构化v2描述(与OpenBrush相同模式);品牌分割为openbrush(15,920件二维艺术)和openartifacts(15,407件摄影对象);媒介标记和归属感知:7,357行有命名艺术家,其中6,311行通过像素签名验证;采用CC0-1.0许可证(公共领域奉献,无保留权利);每行包含完整来源信息:source、landing_page、sha256。数据集结构包括图像字段(图像数据)、宽度和高度(像素尺寸)、品牌(openbrush/openartifacts/ambiguous)、主题(animals)、标签(规范化描述性标签)、9个描述部分(主题、动作、设置、情绪、风格描述、照明、颜色、构图)、完整描述、媒介(如油画、雕塑、金属制品)、艺术家(或unknown)、签名文本(如有)、签名验证(仅当为true时可信)、归属置信度、铭文文本(如有)、铭文可读性和验证性、标题(如有)、权利信息(许可证)、来源和完整性字段。描述生成采用双模型视觉管道:Gemma 4 31B生成完整9部分描述并处理路由,Gemini 3 Flash验证媒介、归属和铭文。标签经过规范化以匹配OpenBrush约定。许可证为CC0 1.0(公共领域奉献),所有图像均来自博物馆和图书馆的开放访问计划,并在摄入时验证(rights_status: safe)。统计数据:总图像31,511张,openbrush 15,920张,openartifacts 15,407张,ambiguous 184张,命名艺术家7,357行,签名验证6,311行,铭文验证12,139行。主要媒介包括雕塑、陶瓷、绘图等;主要标签包括动物、单色、自然等;主要来源包括大都会博物馆开放访问、荷兰国家博物馆OAI等。
OpenArt — Animals & the Natural World is the animals subject collection of the OpenArt family of open, public-domain art datasets: 31,511 works (15,920 paintings/illustrations, 15,407 photographed objects, 184 unclassified), each paired with a structured VLM caption plus medium, attribution and inscription metadata. It covers fauna and the natural world across every medium — birds, beasts, fish and insects painted and drawn, but also carved, cast, glazed and woven into sculpture, ceramics, metalwork and prints, making it the single largest OpenArt subject collection. The dataset draws from the open-access APIs of major museums and libraries (the Metropolitan Museum of Art, Rijksmuseum, Cleveland Museum of Art, Art Institute of Chicago, Smithsonian, Europeana, Wikimedia and others). Unlike the painting-only OpenBrush corpus, each OpenArt subject collection is mixed-medium: it contains both 2-D art (paintings, prints, drawings) and photographs of three-dimensional objects (sculpture, ceramics, metalwork, textiles). A brand column splits every row into openbrush (2-D art), openartifacts (photographed objects), or ambiguous for precise filtering. Key features include: 31,511 public-domain images all passing a rights_status: safe gate; structured v2 captions with 9 semantic sections per image (same schema as OpenBrush); brand split into openbrush (15,920, 2-D art) and openartifacts (15,407, photographed objects); medium-tagged and attribution-aware with 7,357 rows carrying a named artist and 6,311 of those signature-verified against the pixels; CC0-1.0 license (public-domain dedication, no rights reserved); and full provenance per row with source, landing_page, and sha256. The dataset structure includes fields for image, width, height, brand, theme, tags, the 9 caption sections (subject, action, setting, mood, style_description, lighting, color, composition), caption_full, medium, artist, signatures, signature_verified, attribution_confidence, inscription_text, inscription_legible, inscription_verified, title, rights information, and provenance fields. Captions were generated with a two-model vision pipeline: Gemma 4 31B for full captioning and routing, and Gemini 3 Flash for verification of medium, attribution, and inscription. Tags were normalized to match OpenBrush conventions. The license is CC0 1.0, with every image being public domain from museum and library open-access programs and verified at ingest (rights_status: safe). Statistics: total images 31,511, openbrush 15,920, openartifacts 15,407, ambiguous 184, named artist/maker 7,357, signature-verified 6,311, inscription-verified 12,139. Top media include sculpture, ceramic, drawing, etc.; top tags include animal, monochrome, nature, etc.; and top sources include met openaccess, rijksmuseum oai, smithsonian art, etc.




