Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning

Improving Generalized Zero-Shot Learning by Exploring the Diverse Semantics from External Class Names

CVPR 2024 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Generative feature synthesis for zero-shot · 14 citations (OpenAlex)

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 a realistic GZSL setting where unseen classes may be similar or dissimilar to seen classes, and proposes DSECN, which aligns semantics from external ImageNet class names (plus WordNet hyponyms) to visual space via a frozen pretrained classification head.

Problem

Generalized zero-shot learning methods perform poorly when unseen classes are dissimilar to seen classes, since they rely on seen-class semantics and visual features and typically evaluate only on similar unseen classes.

Approach

Introduce diverse semantics from external class names (ImageNet-1K base classes and WordNet hyponym-augmented names) and align them to visual space using the frozen classification head of the pretrained backbone, so the S2V mapping learns a union space that also covers dissimilar unseen classes; the idea can also be plugged into generative and embedding GZSL methods.

Key results

• On CUB with CLIP embeddings under the dissimilar setting, DSECN reaches 47.44 H versus 12.30 for the best prior method (DGZ), a gain of 35.14 points.
• Integrating DSECN into DGZ raises CUB dissimilar-setting H with W2V from 5.67 to 26.34.
• Training takes about 13.9 s on CUB, 37.9 s on AWA2 and 19.7 s on SUN, all under one minute.

Measured effects reported in the paper

Surprising findings

• Existing GZSL methods score very low (often under 10 H) on dissimilar unseen classes, even when they are strong on the similar setting.

Limitations

• Gains for local embedding methods (TransZero) are small because generating local visual feature maps is hard.
• Performance depends on the number of external class names, and the gain depends on what the pretrained classification head covers.
• Relies on a pretrained classification head whose classes overlap with the evaluation datasets, so the dissimilar setting depends on what the pretraining covers.

Experimental setup

Datasets: AWA2 CUB SUN ImageNet-1K ImageNet-21K WordNet

Benchmarks: AWA2 CUB SUN

Metrics: ZSL top-1 accuracy GZSL harmonic mean H

Baselines: ZLA DGZ CN TransZero MSDN HASZSL

Builds on: CN DGZ ZLA TransZero

Closest papers by meaning

• Distinguishing Unseen From Seen for Generalized Zero-Shot Learning (CVPR 2022 · similarity 0.94)
• Data-Free Generalized Zero-Shot Learning (AAAI 2024 · similarity 0.91)
• Data Distribution Distilled Generative Model for Generalized Zero-Shot Recognition (AAAI 2024 · similarity 0.91)
• GenZSL: Generative Zero-Shot Learning Via Inductive Variational Autoencoder (ICML 2025 · similarity 0.90)
• TPR: Topology-Preserving Reservoirs for Generalized Zero-Shot Learning (NeurIPS 2024 · similarity 0.90)
• ZeroDiff: Solidified Visual-semantic Correlation in Zero-Shot Learning (ICLR 2025 · similarity 0.90)