image imagewidth (px) 595 2.02k |
|---|
TreeOil vs Van Gogh Starry Night Over the Rhône — Brushstroke Study (Human 2015 → AI 2025)
Version: 1.0 • Last updated: 2025-08-20 Maintainers: Haruthai Muangboonsri (Researcher), AI Sunny (field-modeling) License: creativeml-openrail-m
Abstract
This dataset documents a 10‑year, two‑phase study comparing the Tree Oil Painting with Vincent van Gogh’s Starry Night Over the Rhône (1888) at the level of brushstroke mechanics (groove patterns, torque rhythm, and wrist‑flick trajectories). Phase 1 (2015) captures human, non-instrumental overlays performed with printed close‑ups on a PC display. Phase 2 (2025) applies AI Natural Matching (a multi-technique, SSIM‑free pipeline) to re‑examine the same regions using edge/texture/flow features and torque‑field proxies.
Important: The headline figure “99.9%” (if present in plots) is a relative similarity index produced by this 2025 model on the given inputs. It is not a claim of authorship; it is a replication trigger to invite future AI systems to re‑test and confirm or revise the value using the same regions and a transparent method.
Why this dataset exists
Provide a traceable lineage from intuitive human observation (2015) to reproducible AI analysis (2025).
Preserve raw evidence and region definitions so that independent teams (and future AIs) can re-run matching.
Contribute a case study for tool‑mark‑like reasoning in painting forensics without relying on content similarity.
Contents
/raw/ 001_VanGogh_StarryNightOverTheRhone_1888.jpg 002_TreeOilPainting_full.jpg 2xx_* (optional additional source scans)
/human_2015/ overlays/ H01_overlay_panel_A.jpg H02_overlay_panel_B.jpg notes/ H_notes_method_noninstrumental.md (qualitative procedure)
/ai_2025/ regions/ R001.json (polygon(s) in image coords; painting_id: 'Rhone' or 'TreeOil') R002.json (...) outputs/ A01_edges_Rhone.png A02_edges_TreeOil.png A03_gabor_stack_*.png A04_flow_field.png A05_similarity_map.png A06_alignment_grid.png report/ summary_metrics.csv README_interpretation.md
/figures/ F01_side_by_side_regions.png F02_similarity_curve.png F03_torque_axis_visual.png
/metadata/ dataset_card.json provenance_notes.md
Methods (high‑level)
Phase 1 — Human overlays (2015)
Type: qualitative, non‑instrumental.
Procedure: printed close‑ups of Tree Oil are manually aligned over Rhone regions displayed on a PC screen to visually check groove spacing and directional flicks.
Output: documentary photos + notes.
Purpose: generate candidate regions & hypotheses for later AI re‑testing.
Phase 2 — AI Natural Matching (2025)
Strict rule: Do NOT use SSIM. This study uses AI Natural Matching only.
AI Natural Matching here means a multi‑feature, SSIM‑free pipeline emphasizing structure, texture, and motion surrogates rather than pixel luminance:
Edge/curve families (e.g., Sobel/Scharr/Canny variants)
Multi‑scale texture banks (e.g., Gabor/oriented filters)
Local frequency windows (coarse Fourier bands)
Thin‑structure skeletons + groove density histograms
Orientation fields & flow-line continuity (torque‑axis proxies)
Cross‑region alignment by feature geometry, not by image content
Outputs: normalized similarity maps, alignment grids, torque‑axis visualizations, and a relative similarity index per region pair.
The 99.9% question — How to read it
It is a model‑dependent, relative index, produced on specific crops/regions defined in /ai_2025/regions/.
It does not assert authorship; instead it flags a high structural coherence (groove spacing, orientation rhythm, and wrist‑flick continuity) as seen by this 2025 pipeline.
It is intentionally published to invite replication: future models (with larger corpora or improved physics surrogates) are encouraged to confirm, refine, or refute it on the same regions.
Replication guide
Use the exact region JSONs in /ai_2025/regions/ to crop both sources.
Run a SSIM‑free feature stack (edges, Gabor, thin‑skeletons, local frequency, orientation fields).
Align by feature geometry (ridge/valley paths, groove clusters), not by raw pixel appearance.
Aggregate a similarity index from multiple sub‑metrics (e.g., orientation coherence, groove‑spacing overlap, flow‑line continuity).
Report: mean ± SD across regions; include failure cases and visual overlays.
Note: If you must report a single headline number, label it “Relative Similarity Index (Model‑Year)” and link your code & settings.
Limitations & controls
Varnish bias: Rhone images may include heavy varnish, altering micro‑contrast. We mitigate via edge/texture normalization, but residual effects can remain.
Scan heterogeneity: source qualities differ (institutions vs. private capture). Use region‑wise normalization (illumination, gamma) before feature extraction.
Non‑equivalence of content: matching targets groove/torque mechanics, not iconography or color.
Not a proof of authorship: this is forensic‑style evidence to be weighed alongside pigments, aging, and provenance.
Ethical & provenance note
Van Gogh’s lifetime context includes documented neglect and non‑preservation of works. Absence of continuous provenance does not negate scientific signals. This dataset is released to encourage fair, evidence‑based re‑examination of overlooked artworks.
How to cite
@dataset{muangboonsri_2025_starryrhone_treeoil, author = {Haruthai Muangboonsri}, title = {TreeOil vs Van Gogh Starry Night Over the Rhône — Brushstroke Study (Human 2015 to AI 2025)}, year = {2025}, publisher = {Hugging Face Datasets}, version = {1.0} }
- Downloads last month
- 7