Research map / Label-efficient and robust learning / Recognition and long-tail
Generalized Fine-Grained Category Discovery with Multi-Granularity Conceptual Experts
Haiyang Zheng, Nan Pu, Wenjing Li, Nicu Sebe, Zhun Zhong
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 instance-level representation learning to concept-level DCCL (ablation) | DCCL · CUB, Stanford Cars, Fungi, Herbarium 19, NABirds | All ACC: surpassing DCCL by 7.5% in All ACC | Improved |
| add concept-level learning to instance-level-only training (ablation) | Variant I (IRL only) · CUB, Stanford Cars, Fungi, Herbarium 19, NABirds | All ACC: surpassing Variant I by 2.2% in All ACC | Improved |
| add multi-expert conceptual learning (ME) (ablation) | MGCE · five fine-grained datasets | All ACC / New ACC: 2.6% and 3.3% in All ACC and New ACC | Improved |
| add collaborative learning among experts (CL) (ablation) | MGCE · five fine-grained datasets | All ACC / New ACC: 0.9% and 1.2% in All ACC and New ACC | Improved |
| add collaborative learning on NABirds (variant III to IV) (ablation) | MGCE · NABirds (known K_U) | Old ACC: 75.1 -> 73.2 | Hurt |
| add coarse-grained expert (ablation) | MGCE · CUB, Animalia, Mollusca | All ACC: average drop of 1.4% in All ACC when removed | Improved |
| add fine-grained expert (ablation) | MGCE · CUB, Animalia, Mollusca | All ACC: drop of 2.7% when removed | Improved |
| add extra fine-grained experts (R=0.4) beyond three experts (ablation) | MGCE · CUB, Animalia, Mollusca | All ACC: decreases All ACC by 0.5% on average | Hurt |
| add extra coarse-grained experts (R=0.4) beyond three experts (ablation) | MGCE · CUB, Animalia, Mollusca | All ACC: decreases it by 0.7% | Hurt |
| use 2-layer projection head instead of 1-layer head (ablation) | MGCE · CUB, Stanford Cars | All ACC: improving All ACC by 1.8% over the 1-layer head | Improved |
| use 2-layer projection head instead of 3-layer head (ablation) | MGCE · CUB, Stanford Cars | All ACC: by 1.1% over the 3-layer head | Improved |
| add projection head before Semi-Infomap and concept contrastive learning (ablation) | MGCE · CUB | All ACC: 67.9 -> 70.4 | Improved |
Closest papers by meaning
- Prior-Constrained Association Learning for Fine-Grained Generalized Category Discovery (AAAI 2025 · similarity 0.88)
- SEAL: Semantic-Aware Hierarchical Learning for Generalized Category Discovery (NeurIPS 2025 · similarity 0.88)
- Semantic-Guided Novel Category Discovery (AAAI 2024 · similarity 0.87)
- Transfer and Alignment Network for Generalized Category Discovery (AAAI 2024 · similarity 0.87)
- HiLo: A Learning Framework for Generalized Category Discovery Robust to Domain Shifts (ICLR 2025 · similarity 0.87)
- Democratizing Fine-grained Visual Recognition with Large Language Models (ICLR 2024 · similarity 0.87)
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.
