Datasetpublic

patryk-bartkowiak/SynGallery-abl0

SynGallery-abl0: fixed-environment baseline Rung 0 of the SynGallery instance-level artwork-recognition ablation ladder. 4,898 MET paintings × 5 camera viewpoints = 24,490 synthetic RGB images at 512×512, paired with their source photos and museum metadata. In this rung, the scene varies nothing — the gallery is one fixed configuration for all 24,490 images, while freezing wall/floor/roof textures + floor material, lighting, glass, frame variant/color, camera pose. Same schema… See the full description on the dataset page: https://huggingface.co/datasets/patryk-bartkowiak/SynGallery-abl0.

sourceHugging Facecc-by-4.0updated 3mo agoView on Hugging Face

0likes122downloads

Dataset Card

SynGallery-abl0: fixed-environment baseline

Rung 0 of the SynGallery instance-level artwork-recognition ablation ladder. 4,898 MET paintings × 5 camera viewpoints = 24,490 synthetic RGB images at 512×512, paired with their source photos and museum metadata.

In this rung, the scene varies nothing — the gallery is one fixed configuration for all 24,490 images, while freezing wall/floor/roof textures + floor material, lighting, glass, frame variant/color, camera pose. Same schema, source images and index↔painting mapping as every other rung — they differ only in which randomization factors are active.

The ablation ladder

Each rung unfreezes one more randomization factor; index maps to the same painting in every rung, so isolating a factor = diffing two adjacent rungs. Three factors vary in every rung by design: the painting image, its scale on the wall, and the placard.

rung	dataset	adds
0	`SynGallery-abl0`	(baseline — fixed environment)
1	`SynGallery-abl1-tex`	wall/floor/roof texture + floor material
2	`SynGallery-abl2-tex-light`	lighting
3	`SynGallery-abl3-tex-light-glass`	glass (p = 0.25)
4	`SynGallery-abl4-tex-light-glass-frame`	frame variant/color
5	`SynGallery`	camera-pose jitter (all variety)

Schema

A single train split, 4,898 rows — one row per scene, bundling everything about that scene: the MET source photo, the five camera views rendered from it, and the artwork's museum metadata.

column	type	meaning
`index`	`int32`	scene index `0..4897`; stable across all SynGallery variants — same index ⇔ same source painting
`met_object_id`	`int32`	MET museum object id of the source painting
`source_image`	`Image`	the MET source photo itself (web-size JPEG)
`image_30` … `image_150`	`Image` ×5	the 512×512 RGBA PNG renders from cameras `30`, `60`, `90`, `120`, `150`
`title`	`string`	artwork title (MET Open Access)
`artist` / `artist_bio`	`string`	artist display name / bio (may be empty)
`object_date`	`string`	display date, e.g. `ca. 1580–85`
`object_begin_date` / `object_end_date`	`int32`	machine-readable date range (null if unknown)
`medium`	`string`	e.g. `Oil on copper`
`classification`	`string`	MET classification, e.g. `Paintings`
`culture` / `department`	`string`	culture (may be empty) / MET curatorial department
`met_url`	`string`	the object's page at metmuseum.org
+ 42 more MET columns	`string` / `bool`	every remaining `MetObjects.csv` field as its own column: `object_number`, `is_highlight` / `is_timeline_work` / `is_public_domain` (bool), `gallery_number`, `accession_year`, `object_name`, `period`, `dynasty`, `reign`, `portfolio`, `constituent_id`, the `artist_*` details (role, nationality, dates, gender, ULAN/Wikidata URLs, …), `dimensions`, `credit_line`, geography (`city` … `river`), `rights_and_reproduction`, `object_wikidata_url`, `metadata_date`, `repository`, `tags` + tag URLs. Sparse fields are empty strings / nulls
`camera_location_30` … `_150`	`float32[3]` ×5	each camera's world position at render time (after per-datapoint jitter)
`camera_rotation_euler_30` … `_150`	`float32[3]` ×5	each camera's world rotation (radians) at render time

Images are embedded bytes only; a render is identified by (index, camera) and the source painting by met_object_id. Rows are ordered by index. There is no held-out split — any row-level split is automatically painting-level, since all 5 views live in the same row.

Scene-randomization factors (wall/floor/roof textures, lighting, frame variant/color, glass presence, painting scale, placard) are baked into the renders but not individually labeled.

Provenance & licensing

Source paintings are photographs of artworks from the MET Open Access collection (CC0). One representative photo per MET object (paintings_train_images_unique.json manifest, consumed in order). Museum metadata is joined from the official MET Open Access `MetObjects.csv` (also CC0) by object id — the complete record, every field as its own column.

The dataset (the rendered images and their packaging) is released under CC BY 4.0 — please attribute the dataset and cite the accompanying paper. The 3D scene bakes in third-party PBR textures and HDRI environment maps; those assets remain under their respective licenses and are distributed here only as non-extractable, baked pixels.

Citation

A paper describing SynGallery is in preparation. Citation details (BibTeX) will be added here once available.

Acknowledgments

This dataset was created in collaboration with ArtiCollect.