Joint Semantic and Latent Attribute Modelling for Cross-Class Transfer Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Cross-class transfer learning addresses vision problems like zero-shot learning and person re-identification.
• Attributes, as shared semantic properties, are crucial for knowledge transfer across object classes.
• Existing methods often overlook undefined, latent visual properties (latent attributes) alongside human-defined semantic attributes.

Purpose of the Study:

• To propose a joint attribute learning model that integrates semantic and latent attributes for improved representation and prediction.
• To extend this model within a multi-task transfer learning framework for unsupervised domain adaptation.
• To enhance performance on challenging vision tasks by leveraging both defined and undefined visual properties.

Main Methods:

• A novel dictionary learning model is proposed, decomposing the dictionary space into semantic, latent discriminative, and latent background attributes.
• The joint attribute learning model is extended using a multi-task transfer learning framework.
• The approach is applied to unsupervised domain adaptation, where target datasets lack labels.

Main Results:

• The proposed linear models are computationally efficient.
• The joint learning of semantic and latent attributes leads to better representation and aids semantic attribute prediction.
• State-of-the-art results were achieved on zero-shot learning and person re-identification tasks.

Conclusions:

• Jointly learning semantic and latent attributes offers significant advantages in computer vision tasks.
• The proposed dictionary learning and multi-task transfer learning framework effectively handles unsupervised domain adaptation.
• The efficient, linear models provide a strong baseline for future research in attribute learning and transfer learning.