Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Zero-Shot Learning—A Comprehensive Evaluation of the Good, the Bad and the Ugly
Yongqin Xian, Christoph H. Lampert, Bernt Schiele, Zeynep Akata
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 |
|---|---|---|---|
| evaluate on the proposed split instead of the standard split | GFZSL · AWA2 ZSL | top-1 accuracy: 79.3 -> 63.8 | Hurt |
| evaluate on the proposed split instead of the standard split | SAE · AWA1 ZSL | top-1 accuracy: 80.6 -> 53.0 | Hurt |
| evaluate on the proposed split instead of the standard split | SYNC · AWA2 ZSL | top-1 accuracy: 71.2 -> 49.3 | Hurt |
| evaluate on the proposed split instead of the standard split | DAP · AWA1 ZSL | top-1 accuracy: 57.1 -> 44.1 | Hurt |
| evaluate on the proposed split instead of the standard split | ALE · aPY ZSL | top-1 accuracy: 30.9 -> 39.7 | Improved |
| tune SSE hyperparameters on validation sets | SSE · AWA1 ZSL | top-1 accuracy: 64.9 -> 76.3 | Hurt |
| tune SSE parameters on the test set | SSE · AWA1 ZSL | top-1 accuracy: 73.8 -> 76.3 | Hurt |
| add novelty detection to CMT | CMT · aPY GZSL | harmonic mean: 2.8 -> 19.0 | Improved |
| train on AWA2 instead of AWA1 (test on same set) | 12 ZSL methods · AWA1/AWA2 cross-dataset | top-1 accuracy: p-value=0.007 | Mixed |
| switch from attribute datasets to ImageNet 20K classes | all ZSL methods · ImageNet ZSL | top-1 accuracy: 0.96 | Hurt |
| use word2vec instead of attributes for a generative model | GFZSL · ImageNet 2H ZSL | top-1 accuracy: 60.6 -> 1.45 | Hurt |
| test on most populated instead of least populated classes | SYNC · ImageNet ZSL 500 classes | top-1 accuracy: 5.83 -> 15.83 | Improved |
Closest papers by meaning
- Zero-Shot Learning - the Good, the Bad and the Ugly (CVPR 2017 · similarity 0.92)
- Recovering the Missing Link: Predicting Class-Attribute Associations for Unsupervised Zero-Shot Learning (CVPR 2016 · similarity 0.91)
- On Zero-Shot Recognition of Generic Objects (CVPR 2019 · similarity 0.91)
- Improved Zero-Shot Classification by Adapting VLMs with Text Descriptions (CVPR 2024 · similarity 0.90)
- Zero-Shot Robustification of Zero-Shot Models (ICLR 2024 · similarity 0.88)
- No "Zero-Shot" Without Exponential Data: Pretraining Concept Frequency Determines Multimodal Model Performance (NeurIPS 2024 · similarity 0.88)
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.
