---
title: 'Picasso Dataset: Scene, Art, & Dark Matter'
url: https://www.emergentmind.com/topics/picasso-dataset
type: topic
---

# Picasso Dataset: Scene, Art, & Dark Matter

The term "Picasso Dataset" refers to three distinct research datasets, each foundational within its respective field: holistic scene reconstruction under physical constraints, computational attribution of artist drawings at the stroke level, and low-mass dark matter direct detection. Each dataset is associated with a separate publication and research agenda, and the name "Picasso" is coincidental across these domains.

## 1. Physics-Constrained Scene Reconstruction: The Picasso Dataset [2602.08058]

The Picasso Dataset, introduced in "Picasso: Holistic Scene Reconstruction with Physics-Constrained Sampling" [2602.08058], is designed for benchmarking the geometric and physics-aware reconstruction of contact-rich real-world tabletop scenes. It provides a testbed for evaluating algorithms that infer object poses and shapes accounting for non-penetration, object interactions, and physical stability, rather than isolated geometric fits.

The dataset comprises 10 manually constructed multi-object indoor scenes, each captured using an Apple iPad Pro with RGB and LiDAR sensors (1920×1440 pixels, ≈60 Hz). The scenes feature between 2 and 8 objects (average ≈5), selected from categories such as plates, bowls, pans, utensils, toys, racks, and household containers, all on approximately planar supports (e.g., tables, sinks, floors).

Each object in every scene is annotated with:
- A high-fidelity mesh or signed-distance representation $S_i$ (acquired via turntable photogrammetry and manual refinement in Blender),
- A similarity transform $T_i = (s_i, R_i, t_i) \in \operatorname{Sim}(3)$ with per-object scale $s_i \in \mathbb{R}^+$, rotation $R_i \in \operatorname{SO}(3)$, and translation $t_i \in \mathbb{R}^3$, all in the calibrated camera frame,
- A contact graph $G = (V, E)$ encoding scene-level object interactions (edges denote contacts; nodes for all objects plus a "0" supporting plane node),
- Per-scene physics-based plausibility metrics assessing non-penetration and mechanical stability.

The acquisition protocol included 10 RGB-D images per scene from varied viewpoints, initial segmentation by SAM2 with exhaustive hand verification, and pose annotation via 2D keypoints and PnP solving, finalized with rendered depth map refinement for accurate geometric alignment.

## 2. Ground Truth, Metrics, and Benchmark Structure

The Picasso Dataset enables digital twins of each real scene to be reconstructed with centimeter-level pose accuracy and explicit contact relationships. Two central physical plausibility metrics are provided:

- **Non-Penetration Score (NPS):** Computes the average penetration depth (in meters) over surface samples, using each object's signed-distance field. Non-penetration is enforced as $\Phi_i(T_i T_j^{-1} b) \geq 0$ for surface sample $b$ between objects $i, j$ (and for contacts with environment/free space).
- **Scene Plausibility Score (SPS):** Determined by the total kinetic energy after a simulated static rollout (20 steps, $\Delta t = 1/240$ s) in PyBullet, assuming canonical mass/inertia. Formally,
  $$
  S^{\mathrm{SPS}} = \sum_{i=1}^N \left( \frac{1}{2} m \|\mathbf{v}_i\|^2 + \frac{1}{2} \boldsymbol{\omega}_i^T J \boldsymbol{\omega}_i \right)
  $$
where $m = 1$, $J = I_{3\times 3}$, $\mathbf{v}_i$ and $\boldsymbol{\omega}_i$ are linear and angular velocities after the rollout. Lower SPS implies higher physical plausibility.

No official train/test split is defined; a 7:3 split by scene is recommended for algorithm development and evaluation. Evaluation metrics include geometric accuracy (ADD-S, ADD-S-AUC), physical plausibility (NPS, SPS), and contact graph accuracy (% of correct predicted edges relative to $G$).

## 3. Representative Scene Examples and Task Relevance

Three scenes illustrate the breadth and annotation quality:

| Scene                      | # Objects | Object Types                                             | Key Contacts / Features         |
|----------------------------|-----------|----------------------------------------------------------|---------------------------------|
| "jenga_2"                  | 6         | Wooden Jenga blocks arranged in a ring                   | Each block touches two neighbors in cycle; $T_i$ pose <1cm error, $S_i$ CAD exact |
| "pan_spatula_steak_broccoli"| 4        | Frying pan, metal spatula, toy steak, plastic broccoli   | Nested point/line contacts; SPS=0.002 J, NPS$\approx$0.004 m (very stable)        |
| "rack_cup_bowl"            | 3         | Cup rack (wireframe), ceramic cup, plastic mixing bowl   | Cup↔rack, bowl↔rack, rack↔table; segmentation and pose challenging                 |

These scenarios test an algorithm’s ability to model inter-object contact, occlusion, thin structures, and multi-category object configurations. The dataset's benchmark tasks are explicitly structured to encourage holistic, physically plausible scene understanding.

## 4. Picasso Dataset for Artist Attribution and Authentication [1711.03536]

A separate, widely cited "Picasso Dataset" comprises 297 digitized line drawings, stroke-segmented and annotated for computational artist attribution and authentication research [1711.03536]. This collection encompasses 80,090 strokes across works by Pablo Picasso (130 drawings; 36,533 strokes), Henri Matisse, Egon Schiele, Amedeo Modigliani, and others. An additional 83 fake drawings (including 24 Picasso forgeries) are included for robustness tests in forgery detection.

Acquisition sources include museum scans, archives, and digital repositories with resolutions ranging from 10–173 pixels/cm. Each drawing is annotated with artist, technique, physical dimensions, digitization metadata, and, at the stroke level, geometric and style-relevant features (e.g., Fourier descriptors, curvature, thickness profiles).

Stroke extraction is performed by skeletonization and graph-based search (bending cost minimization), with iterative longest-path extraction and endpoint/junction updating. Feature extraction combines hand-crafted descriptors and learned representations (e.g., GRU networks on sampled image patches).

The dataset supports supervised classification of individual strokes and entire drawings, achieving ≈80–93% accuracy in artist discrimination and 100% correct detection of fakes in most settings. Access is by request for academic use, with licensing subject to source image rights.

## 5. PICASSO Low-Mass WIMP Search Data [1202.1240]

The acronym "PICASSO" also refers to the dark matter experiment at SNOLAB, which published data on Weakly Interacting Massive Particle (WIMP) interactions with $^{19}$F [1202.1240]. The dataset analyzed arises from 10 ultra-low-background superheated liquid detectors, totaling 0.72 kg of $^{19}$F with 114 kg·d exposure and energy thresholds as low as 1.7 keV.

The primary scientific dataset includes event rates, background rejection parameters (acoustic variables EVAR, FVAR, RVAR), and per-detector mass/exposure and efficiency calibrations. Cross-section upper limits for spin-dependent and spin-independent WIMP-proton interactions are derived by fitting expected rate models to event data, incorporating detector-specific systematic uncertainties.

This dataset is intended for exclusion-curve analysis in the context of low-mass WIMP direct detection. It is not a computer vision benchmark but rather a physical event dataset used for setting constraints on dark matter theory.

## 6. Commonalities, Nomenclature, and Domain-Specific Impact

The three distinct Picasso datasets have no conceptual or methodological overlap beyond nominal similarity. All serve as benchmarks or reference corpora supporting reproducibility and algorithmic comparison. The term "Picasso Dataset" is thus context-dependent, referring either to physics-constrained scene understanding, artist stroke analysis, or rare-event particle astrophysics datasets.

Primary uses include:
- Development and fair evaluation of physically plausible multi-object scene reconstruction (with metric benchmarks for both geometric accuracy and mechanical feasibility),
- Statistical and machine learning-based study of artist attribution, stylistic analysis, and forgery detection at the stroke or drawing level,
- Direct detection dark matter exclusion via threshold event statistics in superheated liquid detectors.

Each dataset has shaped methodological advances in its field and supports ongoing benchmarking and algorithm validation. For domain-specific details, refer to [2602.08058], [1711.03536], and [1202.1240].

Source: https://www.emergentmind.com/topics/picasso-dataset