Research map / Label-efficient and robust learning / Recognition and long-tail
GHOST: Gaussian Hypothesis Open-Set Technique
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 GHOST, a hyperparameter-free post-hoc open-set recognition method that divides maximum logit by a per-class diagonal-Gaussian z-score of the embedding, and evaluates per-class fairness in large-scale OSR.
Problem
Open-set recognition on large-scale ImageNet-1K networks lacks per-class fairness analysis, and existing post-hoc OSR methods (distance-based EVT methods like OpenMax/EVM) require costly hyperparameter/metric selection.
Approach
Model each class's penultimate embedding with a diagonal multivariate Gaussian (per-dimension mean and std from correctly classified training samples), compute the summed absolute z-score of a test embedding against the predicted class, and use the logit divided by this z-score as an open-set score.
Key results
- On MAE-H, GHOST reaches AUROC 0.91 and AUOSCR 0.81 on NINCO versus 0.83/0.78 for MSP, and FPR95 0.47 versus 0.62 for MaxLogit.
- GHOST achieves the lowest coefficient of variation for per-class CCR at 10% FPR across datasets (e.g., 0.21 on NINCO for MAE-H vs 0.45 for MSP).
- Paired t-tests show GHOST's AUROC is significantly better than all compared methods on most datasets (p<1e-6 in the main analysis).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GHOST instead of the strongest listed baseline (MSP, MaxLogit, or NNGuide) | MAE-H OSR · 21K-P Easy, 21K-P Hard, NINCO, OpenImage-O | AUOSCR: .72 -> .75; .69 -> .73; .78 -> .81; .76 -> .84 | Helped |
| Use GHOST instead of the strongest listed baseline | MAE-H OSR · 21K-P Easy, 21K-P Hard, NINCO, OpenImage-O | AUROC: .80 -> .84; .75 -> .81; .83 -> .91; .87 -> .95 | Helped |
| Use GHOST instead of the strongest listed baseline | MAE-H OSR · 21K-P Easy, 21K-P Hard, NINCO, OpenImage-O | FPR95: .63 -> .58; .74 -> .62; .62 -> .47; .49 -> .26 | Helped |
| Use GHOST instead of the strongest listed baseline | MAE-H OSR · 21K-P Easy, 21K-P Hard, NINCO, OpenImage-O | F@C95: 0.53 -> 0.48; 0.64 -> 0.53; 0.51 -> 0.35; 0.39 -> 0.17 | Helped |
| Use GHOST instead of other evaluated methods | MAE-H OSR · 21K-P Easy, 21K-P Hard, NINCO, OpenImage-O at 10% FPR | coefficient of variation: 0.32 vs 0.55; 0.35 vs 0.60; 0.21 vs 0.45; 0.17 vs 0.38 | Helped |
| Use GHOST instead of KNN | MAE-H and ConvNeXtV2-H OSR · 21K-P Easy, NINCO, OpenImage-O | AUOSCR: MAE-H 0.75 vs 0.77, 0.81 vs 0.75, 0.84 vs 0.83; ConvNeXtV2-H 0.74 vs 0.74, 0.79 vs 0.70, 0.83 vs 0.80 | Mixed |
| Use GHOST instead of KNN | MAE-H and ConvNeXtV2-H OSR · evaluated datasets | FPR95 | Hurt |
| Use GHOST instead of the strongest baseline | Swin-T OSR · 21K-P Hard, NINCO, OpenImage-O | AUOSCR: 0.64 -> 0.65; 0.71 -> 0.72; 0.74 -> 0.77 | Helped |
| Use GHOST instead of MaxLogit | DenseNet-121 OSR · 21K-P Easy | AUOSCR: 0.64 -> 0.61 | Hurt |
| Use GHOST instead of MSP | DenseNet-121 OSR · 21K-P Hard | AUOSCR: 0.60 -> 0.58 | Hurt |
Surprising findings
- NNGuide performs far worse than the other methods on the adapted large-scale networks (e.g., AUOSCR 0.49 on NINCO for MAE-H), contrary to its OpenOOD leaderboard standing.
- On DenseNet-121, where 8.27% of per-class feature dimensions reject normality, GHOST's gains shrink, and NNGuide collapses.
Experimental setup
Datasets: ImageNet-1K ImageNet-21K-P Easy ImageNet-21K-P Hard NINCO OpenImage-O Places SUN Textures iNaturalist
Benchmarks: ImageNet-21K-P Easy ImageNet-21K-P Hard NINCO OpenImage-O SUN Places Textures iNaturalist
Metrics: AUOSCR AUROC FPR95 F@C95 coefficient of variation of per-class CCR
Baselines: MSP MaxLogit NNGuide Energy SCALE REACT KNN OpenMax EVM
Builds on: MaxLogit OpenMax NNGuide OSCR
Closest papers by meaning
- Robust Visual Recognition with Class-Imbalanced Open-World Noisy Data (AAAI 2024 · similarity 0.83)
- Frequency Shuffling and Enhancement for Open Set Recognition (AAAI 2024 · similarity 0.83)
- From Coarse to Fine-Grained Open-Set Recognition (CVPR 2024 · similarity 0.83)
- VL-SAM-V2: Open-World Object Detection with General and Specific Query Fusion (NeurIPS 2025 · similarity 0.82)
- Open-World Objectness Modeling Unifies Novel Object Detection (CVPR 2025 · similarity 0.82)
- Learning with Open-world Noisy Data via Class-independent Margin in Dual Representation Space (AAAI 2025 · similarity 0.82)
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.