Research map / Continual learning / Class-incremental
Hierarchical Divide-and-Conquer Grouping for Classification Adaptation of Pre-Trained Models
Ziqian Lu, Yunlong Yu, Qinyue Tong, Jun Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add depth-2 recursive intra-domain grouping to depth-1 HDG-CLIP (ablation) | HDG-CLIP · CUB | Harmonic mean H (%): 74.5 -> 78.2 | Improved |
| Add depth-3 grouping beyond depth 2 (ablation) | HDG-CLIP · AwA2 | Harmonic mean H (%): 94.3 -> 91.9 | Hurt |
| Fine-tune subspace-specific linear layers instead of frozen (ablation) | HDG-CLIP (Depth=1) · SUN | Harmonic mean H (%): 54.0 -> 75.3 | Improved |
| Apply inter-domain grouping (HDG) to frozen CLIP without tuning (ablation) | CLIP · CUB | Harmonic mean H (%): 55.0 -> 63.5 | Improved |
| Apply HDG on top of PSVMA features | PSVMA · SUN | Harmonic mean H (%): 52.3 -> 59.7 | Improved |
| Use similarity-based training-free grouping instead of supervised MLP or Xgboost domain classifier (ablation) | HDG · AwA2 | AccC_N (%): 21.5 -> 98.8 (vs MLP) | Improved |
| Add image-level distance to the domain score (ablation) | HDG · SUN | AccC_N (%): 92.2 -> 94.1 | Improved |
| Add LLM-guided description generation for synthetic novel-class images (ablation) | HDG-CLIP · CUB | Harmonic mean H (%): 6.3% reduction in H without LLMs | Improved |
| Add concept selection filtering of generated samples (ablation) | HDG-CLIP · CUB | Harmonic mean H (%): +1.5% H | Improved |
| Increase number of generated images per class beyond 20 (ablation) | HDG-CLIP · CUB, AwA2, SUN | Harmonic mean H (%) | No clear effect |
| Swap text-to-image generator (GauGAN, GLIDE, DALL-E, SDXL) (ablation) | HDG-CLIP · AwA2 | Harmonic mean H (%) | No clear effect |
| Fine-tune CLIP on AwA2 without grouping, test on CUB | CLIP · AwA2 to CUB | AccH (%): 63.8 -> 17.1 | Hurt |
Closest papers by meaning
- Rethinking Few-Shot Adaptation of Vision-Language Models in Two Stages (CVPR 2025 · similarity 0.87)
- Bayesian Test-Time Adaptation for Vision-Language Models (CVPR 2025 · similarity 0.86)
- Calibrating Multi-modal Representations: A Pursuit of Group Robustness without Annotations (CVPR 2024 · similarity 0.86)
- Test-Time Adaptation of Vision-Language Models for Open-Vocabulary Semantic Segmentation (NeurIPS 2025 · similarity 0.86)
- Preserving Clusters in Prompt Learning for Unsupervised Domain Adaptation (CVPR 2025 · similarity 0.86)
- Leveraging Vision-Language Models for Improving Domain Generalization in Image Classification (CVPR 2024 · 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.
