Research map / Label-efficient and robust learning / Recognition and long-tail
ELTA: An Enhancer against Long-Tail for Aesthetics-oriented Models
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.
Proposes ELTA, a plug-and-play framework (feature-level mixup with label-guided sampling, feature-label similarity alignment, and adaptive distribution sharpening for self-training) that mitigates long-tailed bias in image aesthetics assessment.
Problem
Long-tailed score distributions in image aesthetics assessment (IAA) datasets cause models to regress toward the mean and make large errors on minority (very high/low aesthetic score) samples, hurting generalization and fairness.
Approach
Applies mixup at the feature level with a label-guided sampling strategy to oversample minority aesthetic features, aligns feature cosine-similarity with label similarity, and uses a score-adaptive temperature with grid-searched thresholds to produce reliable pseudo-labels for self-training.
Key results
- On AVA, ELTA reaches PLCC 0.777 and SRCC 0.764, versus 0.770/0.759 for the best prior baseline EAT.
- ELTA attains the lowest low-tail MAE on AVA (0.438 vs 0.490 for EAT) and high-tail MAE (0.426 vs 0.433 for EAT).
- Plugging ELTA modules into other IAA models (e.g., NIMA on AVA) improves PLCC, e.g., NIMA from 0.636 to 0.658 with all modules.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ELTA to the Swin Transformer V2 baseline | Swin Transformer V2 · AVA | PLCC: 0.743 -> 0.777 | Helped |
| Add ELTA to the Swin Transformer V2 baseline | Swin Transformer V2 · AADB | PLCC: 0.740 -> 0.772 | Helped |
| Add ELTA to the Swin Transformer V2 baseline | Swin Transformer V2 · TAD66K | PLCC: 0.507 -> 0.539 | Helped |
| Add ELTA to the Swin Transformer V2 baseline | Swin Transformer V2 · PARA | PLCC: 0.925 -> 0.943 | Helped |
| Add TFA, FLSA, and APDS (ablation) | ELTA on AVA · AVA | MAE (low): 0.616 -> 0.438 | Helped |
| Add TFA, FLSA, and APDS (ablation) | ELTA on AVA · AVA | PLCC: 0.743 -> 0.777 | Helped |
| Add TFA, FLSA, and APDS (ablation) | ELTA on AVA · AVA | SRCC: 0.735 -> 0.764 | Helped |
| Use random sampling instead of the proposed TFA sampling strategy (ablation) | TFA on AVA · AVA | PLCC and SRCC: 4.6% decrease in both PLCC and SRCC | Hurt |
| Restrict mixing to threshold-defined minority features (ablation) | TFA on AVA · AVA | PLCC and SRCC: 0.755 and 0.747 vs. 0.777 and 0.764 | Hurt |
| Remove TFA (ablation) | ELTA on AVA · AVA | MAE (low): increase in MAE by 24.9% for the low segment | Hurt |
| Remove TFA (ablation) | ELTA on AVA · AVA | MAE (high): increase in MAE by 8.7% for the high segment | Hurt |
Surprising findings
- ELTA slightly lags EAT on TAD66K because it lacks theme awareness, and it slightly compromises majority-group performance.
Limitations
- Slight decline in majority-group performance compared with some baselines.
- Slightly lags EAT on TAD66K due to lack of theme awareness.
- Requires a small labeled set and grid search for threshold and sharpening hyperparameters.
Experimental setup
Datasets: AVA AADB TAD66K PARA
Benchmarks: AVA AADB TAD66K PARA
Metrics: PLCC SRCC MAE on low/medium/high percentile groups
Baselines: NIMA HGCN BIAA TANet MaxViT MUSIQ EAT L&FDS RankSim
Builds on: mixup RankSim Deep Imbalanced Regression (LDS/FDS) FixMatch-style pseudo-labeling Swin Transformer V2
Closest papers by meaning
- Advancing Comprehensive Aesthetic Insight with Multi-Scale Text-Guided Self-Supervised Learning (AAAI 2025 · similarity 0.88)
- Bridging Cognitive Gap: Hierarchical Description Learning for Artistic Image Aesthetics Assessment (AAAI 2026 · similarity 0.88)
- ArtiMuse: Fine-Grained Image Aesthetics Assessment with Joint Scoring and Expert-Level Understanding (CVPR 2026 · similarity 0.87)
- Unlocking the Essence of Beauty: Advanced Aesthetic Reasoning with Relative-Absolute Policy Optimization (ICLR 2026 · similarity 0.86)
- AesFA: An Aesthetic Feature-Aware Arbitrary Neural Style Transfer (AAAI 2024 · similarity 0.86)
- Aligning Vision Models with Human Aesthetics in Retrieval: Benchmarks and Algorithms (NeurIPS 2024 · similarity 0.86)
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.