Research map / Label-efficient and robust learning / Domain adaptation and generalization
Learning Fine-grained Domain Generalization via Hyperbolic State Space Hallucination
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 Hyperbolic State Space Hallucination (HSSH), which extends VMamba with style hallucination of state embeddings and a Poincaré-ball consistency loss, for fine-grained domain generalization trained on a single source domain.
Problem
Fine-grained domain generalization: learning fine-grained representations that generalize to unseen target domains under style shifts (illumination, color), when trained only on one source domain.
Approach
Enriches style diversity of Vision Mamba (VMamba) state embeddings by linearly extrapolating per-batch channel mean/std slope and hallucinating features, then projects pre- and post-hallucination embeddings to the Poincaré ball and minimizes their hyperbolic distance (plus a same-coarse-class distance term).
Key results
- HSSH reaches 66.03% average accuracy on CUB-Paintings vs 63.47% for VMamba baseline and 55.10% for FSDG.
- On Birds-31, HSSH averages 90.69% vs 88.24% for VMamba and 82.31% for FSDG.
- On RS-FGDG, HSSH averages 69.65% vs 66.85% for VMamba.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HSSH instead of VMamba | VMamba · CUB-Paintings: C → P | top-1 accuracy: 65.01 -> 67.48 | Helped |
| Use HSSH instead of VMamba | VMamba · CUB-Paintings: P → C | top-1 accuracy: 61.93 -> 64.57 | Helped |
| Use HSSH instead of VMamba | VMamba · RS-FGDG: M → N | top-1 accuracy: +2.46% | Helped |
| Use HSSH instead of VMamba | VMamba · RS-FGDG: N → M | top-1 accuracy: +3.14% | Helped |
| Use HSSH instead of VMamba | VMamba · Birds-31 | top-1 accuracy: 88.24 -> 90.69 | Helped |
| Add SSH (ablation) | VMamba · CUB-Paintings: C → P | top-1 accuracy: 65.01 -> 66.42 | Helped |
| Add SSH (ablation) | VMamba · CUB-Paintings: P → C | top-1 accuracy: 61.93 -> 63.29 | Helped |
| Add HMC (ablation) | VMamba + SSH · CUB-Paintings: C → P | top-1 accuracy: 66.42 -> 67.48 | Helped |
| Add HMC (ablation) | VMamba + SSH · CUB-Paintings: P → C | top-1 accuracy: 63.29 -> 64.57 | Helped |
| Add hallucination at F2 alongside F1 (ablation) | SSH · CUB-Paintings: C → P | top-1 accuracy: 64.78 -> 66.02 | Helped |
| Add hallucination at F3 (ablation) | SSH at F1–F2 · CUB-Paintings: C → P | top-1 accuracy: 66.02 -> 66.81 | Helped |
| Add hallucination at F4 (ablation) | SSH at F1–F3 · CUB-Paintings: C → P | top-1 accuracy: 66.81 -> 67.48 | Helped |
Surprising findings
- Generic DG methods based on ResNet perform far below ImageNet-style baselines and even below some DA methods; VMamba alone already outperforms all prior FGDG methods by a large margin.
Limitations
- Only evaluated on three benchmarks with a single source domain per setting
- Hyperparameters such as curvature c and lambda chosen empirically without broad sensitivity analysis
- Ablation shows gains of about 1-2 points per component, with no variance reported
Experimental setup
Datasets: CUB-200-2011 CUB-200-Paintings Million-AID NWPU-RESISC45 NABirds iNaturalist2017
Benchmarks: CUB-Paintings RS-FGDG Birds-31
Metrics: top-1 accuracy
Baselines: FSDG GMoE SDViT MIRO SAGM RIDG PAN DP-Net VMamba
Builds on: VMamba Vision Mamba MixStyle FSDG
Closest papers by meaning
- DGFamba: Learning Flow Factorized State Space for Visual Domain Generalization (AAAI 2025 · similarity 0.88)
- START: A Generalized State Space Model with Saliency-Driven Token-Aware Transformation (NeurIPS 2024 · similarity 0.87)
- HIDISC: A Hyperbolic Framework for Domain Generalization with Generalized Category Discovery (NeurIPS 2025 · similarity 0.86)
- Hyperbolic Category Discovery (CVPR 2025 · similarity 0.85)
- Exploring Semantic Consistency and Style Diversity for Domain Generalized Semantic Segmentation (AAAI 2025 · similarity 0.85)
- PracticalDG: Perturbation Distillation on Vision-Language Models for Hybrid Domain Generalization (CVPR 2024 · similarity 0.83)
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.