Research map / Label-efficient and robust learning / Domain adaptation and generalization
Generalized Category Discovery under Domain Shift: A Frequency Domain Perspective
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 FREE, a frequency-domain framework for domain-shifted generalized category discovery that separates known/unknown domains by amplitude similarity and applies cross-domain and intra-domain Fourier amplitude perturbations plus difficulty-aware resampling.
Problem
Generalized category discovery degrades when unlabeled data include both unseen categories and unseen domains (domain-shifted GCD); labels for domain membership are unavailable.
Approach
Uses amplitude spectra to separate known- and unknown-domain samples via KNN density against labeled anchors fitted with a 2-component GMM, then performs class-aware low-frequency amplitude swapping (cross-domain) and augmentation-pair amplitude swapping (intra-domain) within a SimGCD-style training loop, with a clustering-difficulty-aware sampler.
Key results
- On DomainNet with Real as known and Painting as unknown, FREE reaches 67.7 All accuracy on Real and 45.6 on Painting, versus 64.4 and 42.1 for HiLo.
- On SSB-C, FREE reports 60.4 All on CUB-C original and 55.7 on corrupted, improving over HiLo's 56.8 and 52.0 respectively.
- Ablations on DomainNet show that domain separation plus domain-specific frequency perturbation improves both known and unknown domains, and difficulty-aware resampling adds further gains on the unknown domain.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FREE instead of HiLo | GCD under domain shift · CUB-C corrupted domain | clustering accuracy (All): 52.0 -> 55.7 | Helped |
| Use FREE instead of HiLo | GCD under domain shift · Scars-C corrupted domain | clustering accuracy (All): 35.6 -> 38.9 | Helped |
| Use FREE instead of HiLo | GCD under domain shift · FGVC-C corrupted domain | clustering accuracy (All): 31.2 -> 35.0 | Helped |
| Add frequency-based domain separation and domain-specific frequency perturbation (ablation) | SimGCD on DomainNet · DomainNet known and unknown domains | clustering accuracy | Helped |
| Add clustering difficulty-aware resampling (ablation) | FREE on DomainNet · DomainNet, particularly the unknown domain | clustering accuracy | Helped |
| Use random frequency perturbation without domain separation (ablation) | GCD under domain shift · DomainNet known and unknown domains | clustering accuracy | Mixed |
| Use class-aware rather than class-agnostic frequency transformations (ablation) | FREE on DomainNet · DomainNet | clustering accuracy | Helped |
Surprising findings
- Randomly swapping amplitude components without domain separation can cause negative transfer, even reducing performance relative to the baseline on unknown domains.
- Existing GCD methods can degrade accuracy on the known domain when unlabeled unknown-domain samples are added.
Limitations
- Restricted to image classification; extension to segmentation and detection is left open.
- Requires access to both known and unknown domain data during training, so it does not handle pure domain generalization.
- Relies on pseudo-labels for the unknown domain, which may be noisy.
Experimental setup
Datasets: DomainNet SSB-C (CUB-C, Scars-C, FGVC-C)
Benchmarks: DomainNet SSB-C
Metrics: clustering accuracy (All, Old, New)
Baselines: RankStats+ UNO+ ORCA GCD SimGCD SPTNet RLCD CDAD-Net HiLo
Builds on: SimGCD FDA (Fourier Domain Adaptation) DINO HiLo
Closest papers by meaning
- HIDISC: A Hyperbolic Framework for Domain Generalization with Generalized Category Discovery (NeurIPS 2025 · similarity 0.91)
- HiLo: A Learning Framework for Generalized Category Discovery Robust to Domain Shifts (ICLR 2025 · similarity 0.91)
- Generalized Category Discovery via Reciprocal Learning and Class-Wise Distribution Regularization (ICML 2025 · similarity 0.90)
- Prior-Constrained Association Learning for Fine-Grained Generalized Category Discovery (AAAI 2025 · similarity 0.90)
- When Domain Generalization meets Generalized Category Discovery: An Adaptive Task-Arithmetic Driven Approach (CVPR 2025 · similarity 0.90)
- 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.