Fine-Grained Feature Generation for Generalized Zero-Shot Video Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Generalized zero-shot video classification (GZSVC) aims to classify videos from seen and unseen classes.
• Existing GZSVC methods often synthesize features for unseen classes using only category names, neglecting richer semantic information.
• Category names alone provide limited relational context, hindering comprehensive video understanding.

Purpose of the Study:

• To propose a fine-grained feature generation model for improved GZSVC.
• To leverage both category names and detailed description texts for more robust feature synthesis.
• To enhance the classification of videos with unseen classes by exploring multi-level semantic information.

Main Methods:

• Developed a fine-grained feature generation model integrating coarse-grained category names and fine-grained description texts.
• Extracted content information from category names and motion information from description texts.
• Subdivided motion into hierarchical constraints, linking events and actions at the feature level.
• Introduced a novel loss function to mitigate positive-negative example imbalance and ensure feature consistency across levels.

Main Results:

• The proposed model demonstrated positive gains in generalized zero-shot video classification.
• Extensive quantitative and qualitative evaluations were conducted on the UCF101 and HMDB51 datasets.
• The framework effectively utilizes richer semantic information beyond simple category names.

Conclusions:

• The fine-grained feature generation model significantly improves GZSVC performance.
• Integrating diverse semantic information (category names and descriptions) is crucial for effective zero-shot learning.
• The proposed hierarchical approach and loss function contribute to robust video classification.