Simple Primitives With Feasibility- and Contextuality-Dependence for Open-World Compositional Zero-Shot Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Open-World Compositional Zero-Shot Learning (OW-CZSL) faces challenges with novel state-object compositions due to large cardinality.
• Conventional methods often ignore the crucial dependencies between states, objects, and their compositions.
• Existing approaches that treat simple primitives independently lead to significant performance degradation.

Purpose of the Study:

• To address the limitations of current OW-CZSL methods by modeling feasibility and contextuality dependencies.
• To improve the recognition of novel state-object compositions in images.
• To develop a robust framework that handles the inherent complexities of compositional learning.

Main Methods:

• Modeling feasibility-dependence (e.g., 'hairy' with 'cat' vs. 'building') and contextuality-dependence (e.g., 'cat' dry vs. wet).
• Designing Semantic Attention (SA) to capture feasibility semantics and reduce impossible predictions.
• Proposing generative Knowledge Disentanglement (KD) to create unbiased image representations and mitigate contextual bias.
• Integrating learned feasibility and contextuality into an independent compositional probability model.

Main Results:

• The proposed SA-and-KD-guided Simple Primitives (SAD-SP) method demonstrates superior or competitive performance.
• Effectively captures and utilizes feasibility and contextuality for improved compositional understanding.
• Achieves state-of-the-art results on three benchmark datasets for OW-CZSL.

Conclusions:

• The developed SAD-SP approach effectively models complex dependencies in OW-CZSL.
• Semantic Attention and Knowledge Disentanglement are key components for enhancing compositional learning.
• The method offers a significant advancement in recognizing novel state-object compositions in diverse image contexts.