Human And Accessories Segmentation Dataset

Home » Case Study » Human And Accessories Segmentation Dataset

Project Overview:

Objective

Creating a versatile dataset for the segmentation of human figures and their accessories is crucial. This dataset plays a vital role in applications such as fashion tech, augmented reality, and retail analytics. To ensure its effectiveness, however, the dataset should encompass a wide range of scenarios.

Scope

Our project focused on collecting a wide range of images that show people in various positions and settings, as well as with different accessories.

Sources

- Models in studio settings, adorned with a plethora of accessories, showcase a diverse array of poses and expressions
- Everyday fashion and accessories are candidly captured in street photography.
- Diverse accessories and attire for special events like weddings and ceremonies.
- Accessories specific to various sports, captured during sports events.
- In professional settings, it’s important to showcase work-related accessories. Moreover, these accessories not only enhance productivity but also reflect personal style. Furthermore, they often include items like sleek laptops, stylish notebooks, and ergonomic desk chairs. Additionally, having the right accessories can create a positive impression during meetings and presentations.

Data Collection Metrics

Studio Shots: 105,000
Candid Street Shots: 95,000
Special Event Photographs: 75,000
Sports Events: 65,000

Annotation Process

Stages

Initial raw data cleaning to eliminate any unsuitable images.
Precise segmentation of the human figures.
Additionally, Detailed segmentation of various accessories like hats, sunglasses, and more.
Finally, A thorough review of annotations for accuracy and completeness.

Annotation Metrics

Total Annotations: 2,400,000 (6 segments per image on average considering a human plus an average of 5 accessories)
Human Segmentation: 400,000
Accessory Segmentations: 2,000,000
Annotations Reviewed: 240,000

Quality Assurance

Stages

Expert Review: A team of fashion and segmentation experts thoroughly reviewed some of the annotations. Additionally, their feedback ensured high-quality results.
Automated Checks: Moreover, algorithms flagged potential errors or overlaps in segmentation, which was crucial for maintaining accuracy.
Inter-annotator Agreement: Furthermore, several annotators were given the same images to ensure segmentation was consistent across the team.

QA Metrics

Annotations Reviewed by Experts: 240,000 (10% of total annotations)
Inconsistencies Identified and Rectified: 24,000 (1% of total annotations)

Conclusion

The Human and Accessories Segmentation Dataset is an innovative collection that meticulously captures the intricate details of human clothing and accessories. Through meticulous collection and comprehensive annotation, this dataset aims to propel advancements in numerous tech applications. Specifically, these applications necessitate a profound understanding of human fashion and accessory trends.