Research map / Label-efficient and robust learning / Recognition and long-tail
Prior-Constrained Association Learning for Fine-Grained Generalized Category Discovery
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 Prior-constrained Association Learning (PAL), a greedy association that uses labeled-class prior to block merges between different known classes, feeding prototypical contrastive learning and joint training with SimGCD for generalized category discovery.
Problem
Generalized category discovery: clustering unlabeled images from both known and novel categories given labeled examples of known categories, where prior association methods use labels only as pre- or post-clustering refinement and thus produce false associations, especially on fine-grained data.
Approach
Incorporate the labeled-class prior directly into every step of a distance-ordered greedy association by tracking group membership and refusing merges that would put two different known classes in one group; the resulting groups serve as prototypes for non-parametric prototypical contrast, which is then combined with a SimGCD parametric classifier in a two-stage schedule.
Key results
- Known-class-number setting: 67.6% All accuracy on CUB and 66.7% on Stanford Cars (parametric eval) and 72.6% CUB / 72.2% Stanford Cars with non-parametric eval, versus 68.2% and 56.9% for CMS on the two datasets, the paper claims +4.4% on CUB and +15.3% on Stanford Cars (per the text; the table's CM
- Unknown-class-number setting with only non-parametric loss: 69.9% All on CUB, 70.5% on Stanford Cars, 56.2% on Aircraft, versus CMS 64.4%, 51.7%, 55.2%.
- Ablation: association alone improves SimGCD on CUB All accuracy from 60.8 to 69.9 and on Stanford Cars from 53.8 to 70.5.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use prior-constrained association instead of association without the prior constraint (ablation) | Association-based non-parametric classifier · Stanford Cars | ACC, All: 65.9 -> 70.5 | Helped |
| Use prior-constrained association instead of association without the prior constraint (ablation) | Association-based non-parametric classifier · Aircraft | ACC, All: 47.5 -> 56.2 | Helped |
| Use prior-constrained association instead of association without the prior constraint (ablation) | Association-based non-parametric classifier · CUB | ACC, All: 67.9 -> 69.9 | Helped |
| Add association-based non-parametric classification to parametric classifier training (ablation) | GCD model · CUB | ACC, All (parametric evaluation): 60.8 -> 67.6 | Helped |
| Add association-based non-parametric classification to parametric classifier training (ablation) | GCD model · Stanford Cars | ACC, All (parametric evaluation): 53.8 -> 66.7 | Helped |
| Add parametric classifier training to association-based non-parametric classification (ablation) | GCD model · CUB | ACC, All (association evaluation): 69.9 -> 72.6 | Helped |
| Add parametric classifier training to association-based non-parametric classification (ablation) | GCD model · Stanford Cars | ACC, All (association evaluation): 70.5 -> 72.2 | Helped |
| Remove the non-parametric-only warm-up stage and train both classifiers in one stage (ablation) | Joint non-parametric and parametric training · CUB | ACC, All (parametric evaluation): 60.8 -> 65.1 | Mixed |
| Remove the non-parametric-only warm-up stage and train both classifiers in one stage (ablation) | Joint non-parametric and parametric training · CUB | ACC, All (association evaluation): 69.9 -> 66.9 | Hurt |
| Reduce association subset size from 100% to 50% (ablation) | Prior-constrained association · CUB | ACC, All: 69.9 -> 60.8 | Hurt |
| Reduce association subset size from 100% to 50% (ablation) | Prior-constrained association · Aircraft | ACC, All: 50.8 -> 56.2 | Helped |
| Set non-parametric loss weight β to 0.1 rather than 0.05 (ablation) | Joint training on Stanford Cars · Stanford Cars | ACC, All (parametric evaluation): 71.9 -> 72.2 | Helped |
Surprising findings
- Adding the prior constraint to Semi-DBSCAN improved it enough to surpass the proposed association on CUB.
- Cutting the association subset size to 30% hurt CUB sharply (69.9 to 45.2 All), while Aircraft and CIFAR-100 benefited from subset association.
Limitations
- Best suited to datasets with a moderate number of images per category; large per-category counts add association noise.
- Pairwise distance computation scales quadratically with dataset size, so large datasets are slow without subset association.
- Parametric classifier requires known ground-truth class number; non-parametric variant estimates class count, which is often underestimated (up to 18% error).
Experimental setup
Datasets: CUB-200 Stanford Cars FGVC-Aircraft Herbarium19 CIFAR-100 ImageNet-100
Benchmarks: CUB-200 Stanford Cars FGVC-Aircraft Herbarium19 CIFAR-100 ImageNet-100
Metrics: clustering accuracy (ACC)
Baselines: k-means RankStats+ UNO+ GPC GCD XCon DCCL PromptCAL PIM μGCD CMS SimGCD
Builds on: SimGCD GCD ProtoNCE DBSCAN
Closest papers by meaning
- Active Generalized Category Discovery (CVPR 2024 · similarity 0.90)
- Generalized Category Discovery via Reciprocal Learning and Class-Wise Distribution Regularization (ICML 2025 · similarity 0.90)
- A Unified Knowledge Transfer Network for Generalized Category Discovery (AAAI 2024 · similarity 0.90)
- Generalized Category Discovery under Domain Shift: A Frequency Domain Perspective (NeurIPS 2025 · similarity 0.90)
- Towards Understanding Parametric Generalized Category Discovery on Graphs (ICML 2025 · similarity 0.89)
- Transfer and Alignment Network for Generalized Category Discovery (AAAI 2024 · similarity 0.89)
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.