Learning multiple relative attributes with humans in the loop.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Semantic attributes offer a more intuitive image annotation method than traditional approaches.
• Existing methods require extensive supervision and high-quality data, limiting their practical use.
• Transfer learning and active learning are strategies to address data limitations.

Purpose of the Study:

• To develop a framework that overcomes limitations of existing image annotation methods.
• To enable learning of multiple relative attributes simultaneously by leveraging their dependencies.
• To integrate human feedback into the learning loop for minimal guidance.

Main Methods:

• A joint active learning to rank framework with pairwise supervision was formulated.
• The framework optimizes multiple ranking functions dependently.
• Pairwise queries, such as "which image is more natural?", were used for human input.

Main Results:

• The proposed method achieved superior retrieval performance compared to state-of-the-art approaches.
• Significantly less human input was required compared to existing methods.
• The framework demonstrated effectiveness on real image datasets.

Conclusions:

• The joint active learning framework effectively addresses limitations in semantic attribute learning for image annotation.
• Simultaneous learning of multiple attributes and active human involvement lead to improved performance and reduced data requirements.
• The method offers a more efficient and scalable solution for image content description.