Relationship-Incremental Scene Graph Generation by a Divide-and-Conquer Pipeline With Feature Adapter.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Scene Graph Generation (SGG) identifies semantic relationships between objects in images.
• Incremental learning presents challenges like old class shift and background shift in SGG.

Purpose of the Study:

• To address the challenges in Relationship-Incremental Scene Graph Generation (RISGG).
• To propose a novel network, DaCFA-Net, for effective RISGG.

Main Methods:

• A Divide-and-Conquer (DaC) pipeline decouples relationship class recognition to mitigate old class shift.
• A Feature Adapter (FA) bridges feature space gaps and mines old relationship information.
• The combined DaC and FA form the DaCFA-Net architecture.

Main Results:

• DaCFA-Net demonstrates significant performance gains on benchmark datasets.
• The proposed method achieves approximately 20% improvement over existing SGG baselines on the VG dataset.

Conclusions:

• DaCFA-Net effectively handles old class and background shifts in RISGG.
• The approach offers a robust solution for incremental learning in scene graph generation.