Research map / Label-efficient and robust learning / Self-supervised and contrastive
AACP: Aesthetics Assessment of Children’s Paintings Based on Self-Supervised Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds an AACP dataset (20k DALL-E-generated unlabeled paintings plus 1.2k expert-annotated children's paintings with eight attributes) and a self-supervised ConvMAE-based model with spatial/channel perception and SVD-based disentangled evaluation for aesthetic assessment.
Problem
Aesthetics assessment of children's paintings is hard due to limited labeled data and the need for multi-attribute evaluation, while existing IAA datasets and single-score image-to-score models are poorly suited to children's paintings.
Approach
Pre-train a simplified ConvMAE encoder with masked auto-encoding on generated children's paintings, then fuse its latent codes into spatial (AdaIN-style) and channel (1D-conv cross-channel) perception modules and map features to eight attributes via SVD-based disentangled evaluation.
Key results
- On the AACP dataset, the full model reaches SRCC 0.61 and LCC 0.65 versus 0.36 SRCC for the best baseline (MLSP), with MSE 0.08.
- In a 15-participant user study on 25 paintings, the method scored 9.7 mean with 0.8 std, the highest mean and lowest std among compared methods.
- Ablation shows removing self-supervised learning drops SRCC from 0.61 to 0.16.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add self-supervised learning (ablation) | Ours · AACP | SRCC: 0.16 -> 0.61 | Helped |
| add self-supervised learning (ablation) | Ours · AACP | LCC: 0.20 -> 0.65 | Helped |
| add self-supervised learning (ablation) | Ours · AACP | MSE: 0.12 -> 0.08 | Helped |
| add disentangled evaluation network (ablation) | Ours · AACP | SRCC: 0.49 -> 0.61 | Helped |
| add disentangled evaluation network (ablation) | Ours · AACP | LCC: 0.48 -> 0.65 | Helped |
| add disentangled evaluation network (ablation) | Ours · AACP | MSE: 0.14 -> 0.08 | Helped |
| add channel perception module (ablation) | Ours · AACP | SRCC: 0.55 -> 0.61 | Helped |
| add channel perception module (ablation) | Ours · AACP | LCC: 0.56 -> 0.65 | Helped |
| add channel perception module (ablation) | Ours · AACP | MSE: 0.11 -> 0.08 | Helped |
| add spatial perception module (ablation) | Ours · AACP | SRCC: 0.41 -> 0.61 | Helped |
| add spatial perception module (ablation) | Ours · AACP | LCC: 0.42 -> 0.65 | Helped |
| add spatial perception module (ablation) | Ours · AACP | MSE: 0.13 -> 0.08 | Helped |
Surprising findings
- A_LAMP achieves the best EMD (0.14), slightly better than the proposed method (0.38).
Limitations
- Dataset is small (1.2k real annotated images) and relies on expert annotation with subjective labels.
- Proposed method does not win on EMD against A_LAMP and PIAA.
- Generated training images from DALL-E may differ from real children's paintings (FID 7.85).
Experimental setup
Datasets: AACP dataset (generated and expert-annotated children's paintings) AVA AADB
Benchmarks: AACP dataset
Metrics: SRCC LCC EMD MSE user study mean score user study standard error
Baselines: NIMA AADB MLSP A_LAMP PIAA
Builds on: ConvMAE masked auto-encoding SE-Net EQGAN-SA SVD
Closest papers by meaning
- ArtiMuse: Fine-Grained Image Aesthetics Assessment with Joint Scoring and Expert-Level Understanding (CVPR 2026 · similarity 0.87)
- APDDv2: Aesthetics of Paintings and Drawings Dataset with Artist Labeled Scores and Comments (NeurIPS 2024 · similarity 0.87)
- Bridging Cognitive Gap: Hierarchical Description Learning for Artistic Image Aesthetics Assessment (AAAI 2026 · similarity 0.86)
- Advancing Comprehensive Aesthetic Insight with Multi-Scale Text-Guided Self-Supervised Learning (AAAI 2025 · similarity 0.85)
- Unlocking the Essence of Beauty: Advanced Aesthetic Reasoning with Relative-Absolute Policy Optimization (ICLR 2026 · similarity 0.83)
- From Crayons to Code: AI-Driven Insights into a Child’s Mental Health Through Drawings (AAAI 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.