CPPE-5: Medical Personal Protective Equipment Dataset.

Area of Science:

• Computer Vision
• Machine Learning
• Medical Informatics

Background:

• Existing datasets like PASCAL VOC, ImageNet, COCO, and OpenImages focus on broad object categories.
• There is a lack of specialized datasets for subordinate categorization of medical personal protective equipment (PPE).
• Current datasets often lack real-life, non-iconic images in complex scenes, limiting practical application.

Purpose of the Study:

• Introduce CPPE-5, a novel dataset for studying subordinate categorization of medical PPE.
• Provide a dataset with images in complex, natural contexts to enhance model generalizability.
• Enable research on PPE recognition in realistic scenarios.

Main Methods:

• Collected real-life images featuring complex scenes with multiple objects.
• Focused on non-iconic imagery to represent practical usage.
• Annotated images with bounding boxes and positive labels for five PPE categories: coveralls, face shields, gloves, masks, and goggles.

Main Results:

• CPPE-5 contains five distinct medical PPE categories.
• The dataset emphasizes complex scenes and non-iconic, real-life images.
• Analysis of baseline and state-of-the-art models for bounding box detection was performed.

Conclusions:

• CPPE-5 addresses the gap in specialized PPE datasets for subordinate categorization.
• The dataset's design supports the development of models for practical, real-world applications.
• Availability of code, data, and trained models facilitates further research and development.