Semantics-Preserving Graph Propagation for Zero-Shot Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Object detection models typically struggle with novel or rare object categories.
• Existing Zero-Shot Object Detection (ZSD) methods often suffer from domain shift issues due to simple mapping-transfer strategies.

Purpose of the Study:

• To propose a novel Semantics-Preserving Graph Propagation model for Zero-Shot Object Detection (ZSD).
• To address the domain shift problem in ZSD and improve the recognition of unseen object categories.

Main Methods:

• Developed a Semantics-Preserving Graph Propagation model utilizing Graph Convolutional Networks (GCN).
• Employed a graph construction module to build category graphs with diverse correlations.
• Implemented two semantics-preserving modules for enhancing category and region representations via multi-step graph propagation.

Main Results:

• The proposed model effectively leverages semantic descriptions and structural knowledge from category graphs.
• Knowledge transfer through graph propagation enhances the generalization capability of the projection function.
• Experiments on popular datasets show favorable performance against state-of-the-art ZSD methods.

Conclusions:

• The Semantics-Preserving Graph Propagation model offers a robust solution to the domain shift problem in ZSD.
• This approach improves the ability to recognize and localize objects from novel concepts.
• The method demonstrates superior performance and generalization capabilities in Zero-Shot Object Detection tasks.