Visual Recognition by Learning From Web Data via Weakly Supervised Domain Generalization.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Real-world visual recognition tasks often involve training data with noisy labels, such as those found on the web.
• Achieving robust classifiers requires addressing both label noise and enhancing generalization to unseen target domains.

Purpose of the Study:

• To propose a weakly supervised domain generalization (WSDG) method for visual recognition tasks using noisy web data.
• To develop a technique that can effectively handle label noise in source domain data and improve generalization to arbitrary target domains.

Main Methods:

• A multi-instance learning (MIL) formulation is employed by partitioning training samples into clusters (bags) and identifying clean samples within each bag.
• The WSDG approach learns a classifier for each category and latent domain, integrating multiple classifiers during testing for improved generalization.
• The method is extended to incorporate textual descriptions as privileged information during training.

Main Results:

• The proposed WSDG method demonstrates effectiveness in learning robust classifiers from noisy web data.
• Experiments on benchmark datasets show significant improvements in generalization capability compared to existing methods.
• The integration of multiple classifiers from latent domains enhances performance in the testing stage.

Conclusions:

• The developed WSDG approach offers a robust solution for visual recognition tasks trained on noisy web data.
• The method successfully addresses the challenges of label noise and domain generalization.
• The findings highlight the potential of WSDG for real-world visual recognition applications.