End-to-End One-Shot Human Parsing.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Existing human parsing methods are restricted to predefined categories, limiting their use in dynamic fashion applications with novel item classes.
• Practical fashion analysis requires parsing humans into an open set of classes, adaptable to unseen categories during testing.

Purpose of the Study:

• To define and address the novel one-shot human parsing (OSHP) task, enabling parsing into an open set of classes.
• To develop an effective model, EOP-Net, capable of handling challenges like small object sizes, testing bias, and similar part discrimination in OSHP.

Main Methods:

• Proposed an End-to-end One-shot human Parsing Network (EOP-Net) for parsing images into coarse-grained and fine-grained human classes.
• Implemented gradual smoothing of training-time prototypes for robust class representations and a dynamic objective for enhanced feature learning and transferability.
• Incorporated a contrastive loss at the prototype level to improve discrimination between visually similar human parts.

Main Results:

• EOP-Net demonstrated strong performance on tailored OSHP benchmarks, outperforming existing one-shot segmentation models significantly.
• The method effectively addresses challenges of small object sizes, mitigates testing bias, and improves discrimination of similar human parts.
• The developed approach shows adaptability to novel classes, crucial for real-world fashion applications.

Conclusions:

• The novel OSHP task and EOP-Net provide a flexible framework for human parsing in scenarios with evolving class definitions.
• EOP-Net establishes a robust baseline for one-shot human parsing, paving the way for future research in adaptable visual recognition systems.
• The proposed methods offer significant improvements for practical applications, particularly in the fashion domain, by enabling parsing into open-ended categories.