Windows Segmentation Dataset

Home » Case Study » Computer Vision » Windows Segmentation Dataset

Project Overview:

Objective

Windows Segmentation Dataset: To construct a comprehensive dataset centered around images of various windows, we aim to propel advancements in AI models specifically for architectural design tools, real estate apps, and virtual staging software.

Scope

The compilation includes images of windows from various architectural styles, time periods, and building types. Each image segments and categorizes the window type, along with providing metadata regarding its style, size, and frame materials.

Sources

After accessing architectural design archives, a careful collection and successful curation of an array of design concepts followed. This process involved meticulously exploring various design archives and selecting the most compelling concepts. These ideas were then organized and curated to ensure they represent the best of architectural design.
Through close collaboration with real estate agencies, we achieved a meticulously collected and thoughtfully curated comprehensive assortment of property-related visuals. Furthermore, active engagement with these agencies ensured a thorough understanding of their needs. Consequently, we compiled a diverse array of visuals, including high-quality photographs and immersive virtual tours. Additionally, this collection was continuously refined through feedback loops, further enhancing its relevance and effectiveness in marketing campaigns.
By incorporating user-submitted photos of homes and commercial buildings into the collection, we have successfully created a diverse compilation. Furthermore, we have actively collected a wide range of images, enriching the content with various architectural styles and settings.
Furthermore, the utilization of drone captures of skyscrapers and tall buildings has contributed to the creation of a thoughtfully collected and visually comprehensive dataset. Additionally, these captures provide detailed insights into the architectural nuances and structural integrity of these towering edifices.
Gathered photos from construction and window manufacturing companies to enrich the collection with a successfully curated range of industry-related visuals.

Data Collection Metrics

Total Window Images: 320,000
Residential Windows: 150,000
Commercial Building Windows: 90,000
Historical Building Windows: 40,000
Modern Design Windows: 30,000
Miscellaneous Windows: 10,000

Annotation Process

Stages

Image Pre-processing: First, we adjust for optimal clarity, lighting, and framing to enhance the image quality. This step ensures that the subsequent processes can be carried out with high accuracy.
Window Segmentation: Then, we label the specific type of window, such as bay, sash, or casement. This categorization helps in distinguishing different window styles and supports better data organization.
Metadata Annotation: In addition, we add details such as architectural style (e.g., Victorian or Modernist), frame material (e.g., wood or aluminum), and the presence of treatments like blinds or curtains. This comprehensive annotation provides valuable context and enriches the dataset.
Validation: Finally, human experts and preliminary window recognition models verify the annotations. This combination ensures the accuracy and reliability of the information, maintaining high standards of quality control.

Annotation Metrics

Total Window Segmentation Annotations: 320,000
Metadata Annotations: 320,000

Quality Assurance

Stages

Automated Window Recognition Verification: Preliminary models validate the segmented windows for consistency.
Peer Review: Furthermore, in the Peer Review stage, a separate set of annotators conducts a second review of the annotations.
Inter-annotator Agreement: Additionally, during the Inter-annotator Agreement process, multiple annotators revisit some images to maintain a high level of consistency in the annotation process.

QA Metrics

Annotations Validated using Window Recognition: 160,000 (50% of total images)
Peer Reviewed Annotations: 96,000 (30% of total images)
Inconsistencies Identified and Rectified: 6,400 (2% of total images)

Conclusion

The Windows Segmentation Dataset is poised to revolutionize the fields of architecture and real estate through AI-driven advancements. This diverse and precise dataset is destined to become a cornerstone in enhancing tools used by architects, designers, and real estate professionals. Consequently, it will lead to more informed and visually stunning projects and presentations.

Let's Discuss your Data collection Requirement With Us

To get a detailed estimation of requirements please reach us.