Learning Graph Embeddings for Open World Compositional Zero-Shot Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Compositional Zero-Shot Learning (CZSL) aims to identify novel combinations of known visual attributes.
• Standard CZSL methods often assume a closed set of possible compositions at test time.
• The open-world setting presents a challenge with an unlimited number of potential unseen compositions.

Purpose of the Study:

• To develop a method for CZSL that operates effectively in an open-world setting.
• To overcome the limitation of pre-defined composition spaces in standard CZSL.
• To improve the recognition of unseen compositions of objects and states.

Main Methods:

• Proposed Compositional Cosine Graph Embeddings (Co-CGE) approach.
• Utilized a graph convolutional neural network to model dependencies between states, objects, and their compositions.
• Incorporated feasibility scores for unseen compositions to refine representations and guide learning.

Main Results:

• Co-CGE achieved state-of-the-art performance on standard CZSL benchmarks.
• Demonstrated superior performance compared to existing methods in the challenging open-world CZSL scenario.
• The graph-based approach effectively propagated information from seen to unseen concepts.

Conclusions:

• Co-CGE successfully addresses the open-world CZSL problem by handling an unlimited compositional space.
• The method's ability to estimate composition feasibility is crucial for robust performance.
• This work advances the field of CZSL by enabling recognition of a wider range of unseen combinations.