Single-person Portrait Matting Dataset

Home » Case Study » Single-person Portrait Matting Dataset

Project Overview:

Objective

To compile a specialized dataset for single-person portrait matting, which supports advancements in digital photography, image editing software, augmented reality, and more, we need to address several key issues.

Scope

This collection showcases single-person portraits in a variety of settings, moods, and lighting conditions. Additionally, the detailed annotations help to clearly distinguish the subject from the background, allowing for accurate image matting.

Sources

Studio portraits with diverse backgrounds and lighting setups.
Outdoor shots across different times of day and locations.
Indoor shots in homes, offices, and other settings.
Varied expressions, poses, and outfits.
Images with different hair types, accessories, and props.

Data Collection Metrics

Total Data Points: 300,000 images
Studio Portraits: 120,000
Outdoor Shots: 100,000
Indoor Casual Settings: 80,000

Annotation Process

Stages

Raw Data Cleaning: First, remove any images that are blurry, have significant obstructions, or are redundant. This ensures that the dataset remains high-quality and useful.
Foreground-Background Segmentation: Next, annotate to clearly distinguish the single person from the surrounding background. This step is crucial for accurate segmentation.
Hair & Fine Details Matting: Then, pay special attention to hair and other intricate details. These areas usually pose challenges in matting, so careful work here is essential.

Annotation Metrics

Total Foreground-Background Segmentations: 300,000
Hair & Fine Detail Annotations: 300,000
Annotations Reviewed: 60,000 (20% of total annotations for quality assurance)

Quality Assurance

Stages

Expert Review: Specialists in digital photography and image processing have carefully analyzed a selection of the dataset. Additionally, they provided detailed feedback on the quality and accuracy of the images.
Automated Consistency Checks: Advanced tools have detected possible inconsistencies or gaps in annotations. Therefore, steps will be taken to address these issues to ensure the dataset’s reliability.
Inter-annotator Agreement: By having multiple annotators review overlapping image sets, standardization was ensured. Moreover, this collaborative approach has significantly improved the consistency of the annotations.

QA Metrics

Annotations Reviewed by Experts: 60,000 (20% of total annotations)
Inconsistencies Identified and Rectified: 12,000 (4% of total annotations)

Conclusion

The Single-person Portrait Matting Dataset is a valuable resource for applications that aim to separate subjects from their backgrounds in portrait images. By focusing on the details of the human figure, especially fine elements like hair, this dataset can significantly improve portrait photography and image editing tasks. As a result, it has the potential to bring about notable advancements in these fields.