Learning to Infer Unseen Single-/ Multi-Attribute-Object Compositions With Graph Networks.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current methods struggle with recognizing complex attribute-object compositions and learning attribute-object relationships.
• Existing models are often limited to single-attribute-object recognition, hindering the understanding of intricate concepts.

Purpose of the Study:

• To develop a flexible model for recognizing both single and multi-attribute-object compositions.
• To enable machines to learn complex relations and transfer knowledge between attributes and objects.
• To improve the accuracy of inferring unseen attribute-object compositions.

Main Methods:

• Proposed an attribute-object semantic association graph model with nodes representing attributes and objects.
• Utilized contrastive loss to minimize misclassifications of similar compositions.
• Introduced a novel balance loss to mitigate domain bias and improve prediction of seen compositions.
• Built a large-scale Multi-Attribute Dataset (MAD) with over 116,000 images and 8,000 categories.
• Developed two new evaluation metrics, Hard and Soft, for multi-attribute scenarios.

Main Results:

• The proposed graph model effectively handles both single- and multi-attribute-object composition recognition.
• Contrastive and balance losses significantly reduced misclassifications and domain bias.
• Experiments on MAD and benchmark datasets demonstrated superior performance compared to existing methods.
• The new metrics provide a comprehensive evaluation framework for multi-attribute recognition.

Conclusions:

• The attribute-object semantic association graph model offers a flexible and effective approach for complex concept learning in machines.
• The developed dataset and metrics advance the field of multi-attribute composition recognition.
• This work paves the way for more robust and generalizable machine understanding of compositional concepts.