Research map / Label-efficient and robust learning / Recognition and long-tail
Fine-grained Classes and How to Find Them
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 FALCON, which discovers fine-grained classes from coarsely labeled data by alternating between training a fine-grained classifier (coarse CE + neighbor consistency + confidence + entropy regularization) and solving a discrete optimization for fine-to-coarse class relations.
Problem
Discovering unknown fine-grained classes within coarsely labeled datasets without any fine-grained supervision, while also inferring the relationships between fine and coarse classes.
Approach
Model coarse predictions as the product of fine-grained predictions and a binary fine-to-coarse relation matrix; alternate between training the fine classifier with coarse CE plus local-consistency and confidence losses and solving a linearized integer quadratic program for the relation matrix (with a balance term), and handle multiple datasets with different coarse taxonomies via dataset-specific
Key results
- FALCON reaches 53.4% Acc / 41.6% ARI on tieredImageNet (608 fine classes), outperforming best baseline ANCOR by 12% in accuracy and 22% in ARI (as stated in text).
- FALCON achieves 59.6% Acc and 42.5% ARI on CIFAR100 versus 51.9% / 36.0% for GEORGE, the best baseline.
- FALCON recovers the ground-truth fine-to-coarse graph exactly (GED 0.0) on Living17, Nonliving26, Entity30, Entity13, CIFAR100 and CIFAR-SI, while SCGM has GED 30.7 to 98.7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FALCON instead of SCAN-C | fine-grained class discovery · Living17 | accuracy: 67.1 -> 71.8 | Helped |
| use FALCON instead of SCAN-C | fine-grained class discovery · Nonliving26 | accuracy: 60.4 -> 65.7 | Helped |
| use FALCON instead of SCAN-C | fine-grained class discovery · Entity30 | accuracy: 60.6 -> 65.1 | Helped |
| use FALCON instead of SCAN-C | fine-grained class discovery · Entity13 | accuracy: 57.7 -> 63.6 | Helped |
| use FALCON instead of GEORGE | fine-grained class discovery · CIFAR100 | accuracy: 51.9 -> 59.6 | Helped |
| use FALCON instead of GEORGE | fine-grained class discovery · CIFAR68 | accuracy: 59.6 -> 60.4 | Helped |
| use FALCON instead of GEORGE | fine-grained class discovery · CIFAR-SI | accuracy: 51.2 -> 55.6 | Helped |
| use FALCON instead of ANCOR | fine-grained class discovery · tieredImageNet | accuracy: 47.8 -> 53.4 | Helped |
| use FALCON instead of ANCOR | fine-grained class discovery · PBMC | accuracy: 44.9 -> 75.8 | Helped |
| add fine-grained consistency and confidence loss L_fine (ablation) | FALCON · CIFAR100 | accuracy: 22.9 -> 59.6 | Helped |
| add entropy regularization L_reg (ablation) | FALCON · CIFAR100 | accuracy: 18.7 -> 59.6 | Helped |
| add coarse supervision L_coarse and dependent confidence loss L_conf (ablation) | FALCON · CIFAR100 | accuracy: 51.0 -> 59.6 | Helped |
Surprising findings
- The coarse cross-entropy loss alone cannot separate fine classes within one coarse class, since all fine classes linked to the same coarse class receive identical gradients.
- Estimating class relations rather than using ground truth does not significantly affect fine-grained accuracy.
Limitations
- Assumes the number of fine-grained classes is known a priori or estimated.
- Assumes samples within a dataset are consistently labeled and each fine class has a single coarse parent.
- Entropy regularization assumes not-severely-imbalanced data; extremely long-tailed distributions need alternative regularizers.
Experimental setup
Datasets: Living17 Nonliving26 Entity30 Entity13 CIFAR100 CIFAR68 CIFAR-SI tieredImageNet PBMC
Benchmarks: Living17 Nonliving26 Entity30 Entity13 CIFAR100 CIFAR68 CIFAR-SI tieredImageNet PBMC
Metrics: clustering accuracy adjusted rand index graph edit distance macro accuracy
Baselines: SCAN SCAN-C ANCOR SNCA GEORGE SCGM ERM
Builds on: SCAN MoCoV3 SCGM ANCOR
Closest papers by meaning
- Novel Class Discovery for Ultra-Fine-Grained Visual Categorization (CVPR 2024 · similarity 0.87)
- Advancing Fine-Grained Classification by Structure and Subject Preserving Augmentation (NeurIPS 2024 · similarity 0.87)
- Navigating Real-World Partial Label Learning: Unveiling Fine-Grained Images with Attributes (AAAI 2024 · similarity 0.86)
- Twofold Debiasing Enhances Fine-Grained Learning with Coarse Labels (AAAI 2025 · similarity 0.86)
- From Coarse to Fine-Grained Open-Set Recognition (CVPR 2024 · similarity 0.85)
- Few-Shot Fine-Grained Image Classification with Progressively Feature Refinement and Continuous Relationship Modeling (AAAI 2025 · similarity 0.85)
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.