Research map / Label-efficient and robust learning / Recognition and long-tail
Novel Class Discovery for Ultra-Fine-Grained Visual Categorization
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.
Introduces the UFG-NCD task (novel class discovery for ultra-fine-grained soybean leaf categorization) and proposes RAPL, combining channel-wise region alignment with semi-supervised proxy learning on SoyAgeing datasets.
Problem
Ultra-fine-grained visual categorization requires expensive annotation, and existing novel class discovery methods fail to capture the subtle inter-class differences among unlabeled ultra-fine-grained classes.
Approach
A channel-wise region alignment module assigns feature channels to spatial regions, and a semi-supervised proxy learning scheme uses class-level proxies for proxy-guided supervised classification, proxy regularization, and proxy-guided contrastive learning to transfer knowledge from labeled to unlabeled classes.
Key results
- On SoyAgeing-R1 test split (task-agnostic), RAPL reaches 58.99% All accuracy versus 52.33% for rKD and 53.24% for IIC.
- RAPL improves over the best baseline by a minimum of 3.94% (SoyAgeing-R6) and a maximum of 8.38% (SoyAgeing-R5) on All classes.
- Under the GCD setting, RAPL outperforms GCD and XCon by over 10% on SoyAgeing-R1, R3, R4, R5 and by 6.22% on SoyAgeing-R6.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use proxy-guided contrastive learning instead of vanilla contrastive learning (ablation) | RAPL · SoyAgeing-R1 | accuracy: 56.36 -> 58.99 | Helped |
| Add channel-wise region alignment (ablation) | RAPL · SoyAgeing-R1 | accuracy: 56.97 -> 58.99 | Helped |
| Add proxy regularization (ablation) | RAPL · SoyAgeing-R1 | accuracy: 58.48 -> 58.99 | Helped |
| Use RAPL instead of the strongest baseline | UFG-NCD · SoyAgeing-R5 test, all classes | accuracy: 52.63 -> 61.01 | Helped |
| Use RAPL instead of the strongest baseline | UFG-NCD · SoyAgeing-R4 unlabeled train split | accuracy: 35.71 -> 49.90 | Helped |
Surprising findings
- rKD achieves higher accuracy on Old classes but underperforms on New classes, attributed to its distillation preventing learning of new categories.
Limitations
- Evaluated only on soybean leaf datasets from one source (SoyAgeing), limiting generality to other ultra-fine-grained domains.
- Number of new classes is assumed known a priori.
- Performance drops on SoyAgeing-R6 due to large intra-class variation in the final growth stage.
Experimental setup
Datasets: SoyAgeing-R1 SoyAgeing-R3 SoyAgeing-R4 SoyAgeing-R5 SoyAgeing-R6
Benchmarks: SoyAgeing-R1 SoyAgeing-R3 SoyAgeing-R4 SoyAgeing-R5 SoyAgeing-R6
Metrics: clustering accuracy (ACC) classification accuracy
Baselines: RankStats IL UNO ComEx IIC rKD GCD XCon
Builds on: UNO RankStats rKD XCon MaskCOV
Closest papers by meaning
- Advancing Fine-Grained Classification by Structure and Subject Preserving Augmentation (NeurIPS 2024 · similarity 0.88)
- From Few-way to Many-way: Rethinking Few-shot Fine-grained Image Classification (CVPR 2026 · similarity 0.87)
- Fine-grained Classes and How to Find Them (ICML 2024 · similarity 0.87)
- Democratizing Fine-grained Visual Recognition with Large Language Models (ICLR 2024 · similarity 0.86)
- Navigating Real-World Partial Label Learning: Unveiling Fine-Grained Images with Attributes (AAAI 2024 · similarity 0.85)
- Fine-R1: Make Multi-modal LLMs Excel in Fine-Grained Visual Recognition by Chain-of-Thought Reasoning (ICLR 2026 · similarity 0.84)
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.