Graphonomy: Universal Image Parsing via Graph Reasoning and Transfer.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Existing image parsing models are domain-specific and require extensive retraining for new scenarios.
• Unifying diverse label annotations and granularities into a single model is a significant challenge.
• Lack of methods to discover semantic structures across different label granularities and mine cross-task label correlations.

Purpose of the Study:

• To develop a universal image parsing model adaptable to various domains and label granularities.
• To address the challenges of learning semantic structures and label correlations across different annotation levels.
• To propose a framework that incorporates human knowledge and label taxonomy into graph representation learning.

Main Methods:

• Proposed Graphonomy, a graph reasoning and transfer learning framework.
• Incorporated human knowledge and label taxonomy into intermediate graph representations.
• Utilized two iterated modules: Intra-Graph Reasoning for within-domain semantic graph extraction and Inter-Graph Transfer for cross-domain knowledge sharing.
• Learned global and structured semantic coherency via semantic-aware graph reasoning and transfer.

Main Results:

• Successfully applied Graphonomy to human parsing and panoptic segmentation tasks.
• Demonstrated superior performance compared to state-of-the-art approaches using a standard pipeline.
• Showcased the framework's ability to handle different domains and label granularities effectively.
• Achieved the generation of human parsing at various granularity levels by unifying annotations.

Conclusions:

• Graphonomy offers a robust solution for building universal image parsing models.
• The framework facilitates effective knowledge transfer and learning across diverse computer vision domains.
• Graphonomy enhances feature representation learning through graph-based information propagation and cross-domain collaboration.
• The proposed approach generalizes well and offers additional benefits like multi-granularity parsing.