RichArt

RichArt is a benchmark dataset of 7,087 paintings with 20,882 region-level annotations pairing bounding boxes with rich symbolic narratives extracted from museum catalogs. It is introduced in "Fine-Grained Cross-Modal Retrieval in Art via Region-Level Grounding of Symbolic Narratives" (ICMR 2026).

Dataset Structure

├── annotations/
│   ├── train.json   
│   ├── val.json     
│   └── test.json    
└── data/
    ├── train/ (80% — 5,669 paintings)
    ├── val/   (10% —   709 paintings)
    └── test/  (10% —   709 paintings)

Statistics

General Aspects

Metric	Value
Paintings	7,087
Artists	2,288
Annotated objects	20,882
Unique object labels	6,233
Century range	14th – 21st

Source Distribution:

• 76.5% Web Gallery of Art
• 12.9% Metropolitan Museum of Art
• 10.6% WikiArt

Annotation Granularity

Object Density: Median of 2 objects per painting (max 8).

Narrative Depth: Region descriptions average 35 words, while painting-level descriptions range from 26 to 8,573 words (Mean: 212, Median: 123).

Taxonomy & Style

Categorization: 81.89% of paintings are classified into 7 coarse types (e.g., religious, portrait, landscape) and 25 fine-grained subcategories.

Artistic Style: 40.77% of the corpus is tagged with one of 33 styles (e.g., Baroque, Realism, Tenebrism).

Visual Examples

Structured Painting Metadata

Example of a painting and its metadata (including text highlights of a referring expression and attached description)

Region-Level Symbolic Annotation

Region-level annotation example showing bounding boxes with symbolic narratives on a Dalí painting

Diversity of Annotated Object Regions

Grid of diverse annotated object crops from across the RichArt dataset

Annotation Format

The annotations are provided in JSON format. Each entry represents a single painting and contains high-level metadata alongside region-level "symbolic narratives" grounded in the image.

Schema Definition

Field	Type	Description
id	int	Unique identifier for the painting.
title	string	The official title of the artwork.
artist	string	The name of the artist.
year	int	The year of creation.
coarse_type	string | null	High-level thematic category (e.g., religious, portrait).
first_fine_grained_type	string | null	Primary specific subcategory within the coarse type.
second_fine_grained_type	string | null	Secondary subcategory, if applicable.
first_style	string | null	Primary artistic style (e.g., baroque, early renaissance).
second_style	string | null	Secondary artistic style, if applicable.
description	string	The comprehensive narrative text for the painting from the source museum catalog.
source	string	The origin of the data (e.g., met, wga, or wikiart).
height	int	Height of the digital image in pixels.
width	int	Width of the digital image in pixels.
objects	dict	The core annotation object. Map of symbolic entity names to their groundings.
objects[key].description	string	The specific narrative extracted from the painting description tied to the identified region.
objects[key].bounding_boxes	list	A list of groundings: [score, [x1, y1, x2, y2]].

Generic Example

The following example illustrates a typical entry. Note that a single symbolic entity can be linked to multiple bounding boxes if that element appears in different parts of the painting.

{
    "id": 123,
    "title": "Example Artwork Title",
    "artist": "Artist Name",
    "year": 1500,
    "coarse_type": "religious",
    "first_fine_grained_type": "altarpiece",
    "second_fine_grained_type": null,
    "first_style": "renaissance",
    "second_style": "manierism",
    "description": "A comprehensive description of the entire painting, including historical context...",
    "source": "met",
    "height": 1200,
    "width": 900,
    "objects": {
        "symbolic_entity_name": {
            "description": "A specific narrative excerpt describing this particular element.",
            "bounding_boxes": [
                [
                    0.85, 
                    [100.5, 200.0, 300.2, 450.7]
                ],
                [
                    0.72, 
                    [500.0, 210.5, 650.3, 400.1]
                ]
            ]
        }
    }
}

Key Considerations

ID Value: Some paintings have IDs with values larger than 7,087 because the IDs from the initial set of 12,078 paintings, which have been filtered, have been kept.

Symbolic Narratives: Unlike standard object detection (literal labels like "person"), RichArt uses descriptive phrases (e.g., "the nervous behavior of the hands") to enable fine-grained retrieval.

Coordinates: Bounding boxes use absolute pixel format: [x_min, y_min, x_max, y_max].

Confidence Scores: The first value in the bounding box array represents an alignment score from the grounding process provided by GroundingDINO.

Methodology & Creation

Detailed information regarding the data collection pipeline, annotation methodology, and region grounding verification can be found in the associated research paper and in the corresponding repository.