Part-Object Progressive Refinement Network for Zero-Shot Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Zero-shot learning (ZSL) aims to recognize unseen classes by leveraging semantic information from seen classes.
• Existing ZSL methods often align global visual features or simply combine local part features, neglecting part-object interactions.
• This limitation hinders discriminative and representative knowledge transfer crucial for ZSL.

Purpose of the Study:

• To propose a novel Part-Object Progressive Refinement Network (POPRNet) for improved ZSL.
• To enhance the transferability of semantic knowledge by modeling part-object cooperation.
• To boost the discriminability and recognition performance in ZSL tasks.

Main Methods:

• POPRNet progressively refines semantics through the interplay of object parts and the whole object.
• It incorporates discriminative part semantics and object-centric semantics guided by semantic intensity.
• A semantic-augment transformer (SaT) models part-object relations, and a prototype updating module enhances category prototypes.

Main Results:

• The proposed POPRNet method demonstrates superior and competitive performance on three public benchmark datasets.
• The network effectively refines semantics, improving cross-domain transferability.
• Experimental results validate the effectiveness of the part-object learning approach.

Conclusions:

• POPRNet offers a significant advancement in zero-shot learning by effectively modeling part-object interactions.
• The proposed semantic-augment transformer and prototype updating module contribute to enhanced discriminability and transferability.
• The method provides a more robust approach for recognizing unseen visual categories.