Characters Relationship Segmentation Dataset

Home » Case Study » Characters Relationship Segmentation Dataset

Project Overview:

Objective

Our goal was to create a dataset that focuses on separating the relationships between characters in images. This project aims to advance computer vision applications, particularly in social interaction analysis, storytelling, and human-computer interaction.

Scope

We embarked on a mission to create a comprehensive repository of images showing different social interactions. Each image in our dataset includes detailed annotations that clearly outline the relationships between characters. We carefully collected and annotated these images.

Sources

Photography Enthusiasts: Collaborate with photographers and individuals to get images showing various social interactions, like conversations, handshakes, or group gatherings. By working closely with photographers, we can ensure the authenticity and quality of the images.
Public Image Repositories: Extract images from public image repositories, ensuring they cover diverse content and contexts. This approach allows us to access a wide range of images quickly, providing a rich source of varied social interactions.
Public Submissions: Develop a platform for people to contribute images of social interactions, promoting inclusivity in the relationships depicted. This platform will encourage public participation and ensure a broader representation of social diversity.

Data Collection Metrics

Total Images Collected: 250,000 images (Random Volume Addition)
Photography Enthusiast Contributions: 120,000
Public Image Repository Extracts: 90,000

Annotation Process

Stages

Relationship Segmentation: We annotated each image to define the relationships between characters.
Interaction Type Categorization: Each image was labeled with the type of social interaction it depicted.
Context Description: A brief description of the interaction’s context was provided where applicable.

Annotation Metrics

Images with Relationship Segmentations: 250,000
Interaction Type Labels: 250,000
Context Descriptions: 220,000

Quality Assurance

Stages

Segmentation Accuracy Checks: Employ computer vision algorithms and human reviewers to validate the accuracy of relationship segmentations.
Metadata Validation: Engage social scientists or domain experts to review interaction type categorizations and context descriptions.
Privacy Safeguards: Ensure that publicly contributed images do not inadvertently contain sensitive information, personal identifiers, or private contexts. Implement mechanisms for data contributors to request removal or modifications to their submissions.

QA Metrics

Segmentation Validation Cases: 20,000 (10% of total)
Metadata Authentication Reviews: 30,000 (15% random sampling)
Privacy Audits: 30,000 (for public submissions)

Conclusion

The Characters Relationship Segmentation Dataset is a testament to our commitment to providing high-quality data for AI and machine learning applications. Our dataset offers unparalleled insights into social interactions and human behavior, making it an invaluable resource for research and development across various fields.