Research map / Label-efficient and robust learning / Recognition and long-tail
Navigating Real-World Partial Label Learning: Unveiling Fine-Grained Images with Attributes
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 SoDisam, a partial label learning framework for fine-grained images that uses shared candidate-set attributes as supervision, distinct attributes via attention, and a dynamic disambiguation module.
Problem
Partial label learning is usually evaluated on generic datasets like CIFAR, where candidate labels are easy to tell apart; realistic PLL on fine-grained images with attributes is rarely studied.
Approach
Treat attributes shared by all candidate labels as certain supervision for an attribute-space visual representation, use distinct attributes to guide holistic features via attribute attention, and dynamically prune/reintroduce candidate labels and train on easy samples with CE loss.
Key results
- SoDisam reaches 77.47% top-1 on CUB at ambiguity q=0.01, outperforming the best baseline by up to 2.46 points.
- On AWA, SoDisam reaches 92.66% at q=0.05 and only drops 1.46 points as q rises to 0.2.
- On SUN at q=0.02 and 0.03, SoDisam beats the best baseline ParSE by 2.75 and 3.35 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SoDisam instead of ParSE | CUB (q=0.01) · CUB (q=0.01) | top-1 accuracy: 75.01 ± 0.09% -> 77.47 ± 0.19% | Helped |
| Use SoDisam instead of PiCO | AWA (q=0.05) · AWA (q=0.05) | top-1 accuracy: 89.78 ± 0.46% -> 92.66 ± 0.08% | Helped |
| Use SoDisam instead of ParSE | SUN (q=0.005) · SUN (q=0.005) | top-1 accuracy: 60.90 ± 0.82% -> 63.46 ± 0.19% | Helped |
| Add attribute loss Latt to Lcls (ablation) | SoDisam · CUB (q=0.01), AWA (q=0.05), SUN (q=0.005) | top-1 accuracy: 74.16 -> 73.83; 89.64 -> 89.76; 54.87 -> 54.56 | Mixed |
| Add attribute attention mechanism (ablation) | Lcls + Latt ablation · CUB (q=0.01), AWA (q=0.05), SUN (q=0.005) | top-1 accuracy: 73.83 -> 75.62; 89.76 -> 91.29; 54.56 -> 58.45 | Helped |
| Add easy-sample cross-entropy loss Lce (ablation) | SoDisam without Lce · CUB (q=0.01), AWA (q=0.05), SUN (q=0.005) | top-1 accuracy: 75.62 -> 77.02; 91.29 -> 92.93; 58.45 -> 61.03 | Helped |
| Add dynamic disambiguation module (ablation) | SoDisam without DDM · CUB (q=0.01), AWA (q=0.05), SUN (q=0.005) | top-1 accuracy: 77.02 -> 77.47; 92.93 -> 92.66; 61.03 -> 63.46 | Mixed |
Surprising findings
- Several deep PLL methods that perform near supervised learning on CIFAR-100 perform no better than the simple PRODEN baseline on fine-grained CUB.
Limitations
- Relies on class-level attribute annotations being available for the dataset.
- Attribute-based loss weight alpha degrades performance as it increases, indicating attribute labels can be unreliable for classification.
Experimental setup
Datasets: CUB AWA2 SUN CIFAR-100
Benchmarks: CUB AWA2 SUN
Metrics: top-1 accuracy
Baselines: EXP PRODEN PiCO DPLL ParSE PaPi
Builds on: PRODEN PiCO EXP loss dual attention network (channel attention)
Closest papers by meaning
- Distilling Reliable Knowledge for Instance-Dependent Partial Label Learning (AAAI 2024 · similarity 0.89)
- Realistic Evaluation of Deep Partial-Label Learning Algorithms (ICLR 2025 · similarity 0.89)
- CFDM: Contrastive Fusion and Disambiguation for Multi-View Partial-Label Learning (AAAI 2025 · similarity 0.88)
- Partial Label Learning with a Partner (AAAI 2024 · similarity 0.88)
- Disentangled Partial Label Learning (AAAI 2024 · similarity 0.87)
- Partial Label Causal Representation Learning for Instance-Dependent Supervision and Domain Generalization (AAAI 2025 · similarity 0.87)
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.