Research map / Label-efficient and robust learning / Recognition and long-tail

Learning State-Based Node Representations from a Class Hierarchy for Fine-Grained Open-Set Detection

ICML 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Correspondence and causal representation learning

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 State-FGOD, which encodes a class hierarchy as legal global states over nodes (with openset child nodes) and trains a ViT-B/16 prompt-tuned classifier to detect fine-grained openset samples, with theory showing local/path-probability methods are suboptimal.

Problem

Fine-grained openset detection (FGOD): identifying openset samples that are semantically similar to closed-set classes and assigning them to the correct openset parent node in a class hierarchy, where local top-down detectors accumulate errors and global detectors ignore hierarchical dependencies.

Approach

Model the hierarchy (augmented with openset nodes) as a graph with subsumption and exclusion edges, define legal global states (each corresponding to a leaf or none), and compute state and node probabilities from logits so that the cross-entropy loss updates all ground-truth ancestors and non-ground-truth nodes; train with random removal of nodes to create openset cases.

Key results

• On CUB, State-FGOD reaches AUC 51.04 and OA@50 64.47 with ViT-B/16 prompt tuning, improving AUC by +5.45 over the best baseline.
• On Tiny ImageNet and AWA2, State-FGOD improves AUC by +2.45 and +2.82 over the strongest baseline respectively.
• Applying state formulation to DARTS, Relabel and LOO improves their AUC by 4.89, 4.36 and 7.78 points on CUB.

Measured effects reported in the paper

Surprising findings

• Local detectors with path probability give only marginal gains and can even reduce closed-set accuracy for MSP and HC-EN, showing path probability alone is insufficient.

Limitations

• Evaluated only on image datasets with a single ViT-B/16 backbone and prompt tuning.
• Global state matrix grows with the number of leaf and openset nodes, so scalability to very deep or large hierarchies is not analyzed.
• Gains on Tiny ImageNet and AWA2 are modest with large standard deviations in some settings.

Experimental setup

Datasets: Tiny ImageNet CUB-200-2011 AWA2 Stanford Cars CIFAR-100

Benchmarks: CUB-200-2011 Tiny ImageNet AWA2 Stanford Cars CIFAR-100

Metrics: openset accuracy closed-set accuracy AUC OA@50 HMean hierarchical precision hierarchical recall top-2 accuracy

Baselines: TD MSP HC-EN DARTS Relabel Evidential LOO TD+LOO E-HND

Builds on: Leave-One-Out openset training (Lee et al., 2018) E-HND visual prompt tuning (Jia et al., 2022) Relabel (Lee et al., 2018)

Closest papers by meaning

• From Coarse to Fine-Grained Open-Set Recognition (CVPR 2024 · similarity 0.87)
• The Devil is in the Fine-Grained Details: Evaluating Open-Vocabulary Object Detectors for Fine-Grained Understanding (CVPR 2024 · similarity 0.86)
• Hyp-OW: Exploiting Hierarchical Structure Learning with Hyperbolic Distance Enhances Open World Object Detection (AAAI 2024 · similarity 0.86)
• Cross-domain Open-world Discovery (ICML 2024 · similarity 0.84)
• GUIDED: Granular Understanding via Identification, Detection, and Discrimination for Fine-Grained Open-Vocabulary Object Detection (NeurIPS 2025 · similarity 0.84)
• Towards Evidential and Class Separable Open Set Object Detection (AAAI 2024 · similarity 0.84)