Label-Embedding for Image Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Attribute-based image classification relies on intermediate representations for parameter sharing, crucial for limited training data.
• Existing methods often struggle with data scarcity in classification tasks.

Purpose of the Study:

• To propose a novel framework for attribute-based image classification using label embedding.
• To enhance zero-shot learning performance by treating classification as a label-embedding problem.

Main Methods:

• Viewed attribute-based image classification as a label-embedding problem, embedding each class in an attribute vector space.
• Introduced a compatibility function to measure image-label embedding alignment, learned from labeled samples.
• Evaluated the framework on Animals With Attributes and Caltech-UCSD-Birds datasets.

Main Results:

• The proposed label embedding framework significantly outperformed the Direct Attribute Prediction baseline in zero-shot learning.
• Demonstrated superior performance on benchmark datasets like Animals With Attributes and Caltech-UCSD-Birds.

Conclusions:

• Label embedding provides a robust approach for attribute-based image classification, especially in low-data regimes.
• The framework supports integration of diverse information sources (e.g., hierarchies, text) and spans from zero-shot to standard learning settings.