End-to-end annotation for real estate computer vision pipelines
USA
Model of cooperation:
4 Dedicated Part-time Annotators
Successfully annotated
7,700
floor plans and
10,009
property website screenshots.
Cooperation Background
A US-based PropTech and real estate analytics company focused on building multiple computer vision models for property understanding and classification. Their pipelines relied on annotated datasets covering both architectural floor plans and real estate website screenshots.
The client worked with rapidly evolving models and needed a reliable annotation partner to support continuous dataset creation, iteration, and model improvement.
Client’s Problem
The client was developing multiple small-scale computer vision models, each requiring its own labeled dataset of around 10,000 images. Their internal team handled initial annotation work but needed external support to scale operations and maintain consistency across datasets.
The work included both structured architectural data and highly unstructured web images, which created significant variability in annotation rules.
- Key requirements included:
- Polygon segmentation for floor plans and blueprints
- Image sorting and dataset structuring
- Annotation of real estate website screenshots
- Support for iterative model training cycles
- Scalable team for ongoing annotation needs
In total, the project involved 7,700 architectural floor plans and 10,009 property website screenshots.
Project Challenge
The main challenge was the high variability of real estate website layouts. Each website followed different design patterns, making it difficult to consistently identify elements such as banners, contact forms, cookie notices, and promotional content.
Another challenge was maintaining consistent annotation logic across two very different data types: structured architectural drawings and unstructured web interfaces.
- Additional complexity came from:
- Ambiguous classification of promotional and informational elements
- Distinguishing concessions from standard UI components
- Ensuring consistency across more than 17,700 images
- Supporting iterative model feedback and continuous dataset updates
Mobilunity-BPO Solution
Mobilunity-BPO built a flexible annotation workflow designed to support fast-changing computer vision requirements and multiple concurrent datasets.
The process included
- Structuring the annotation workflow for different data types (floor plans and web screenshots)
- Training a dedicated team of four part-time annotation specialists
- Setting up CVAT-based pipelines for segmentation and detection tasks
- Implementing clear labeling guidelines for complex UI and architectural elements
To support iterative ML development, Mobilunity-BPO also introduced a model-driven quality approach, where feedback from model outputs was used to refine annotation accuracy over time.
The managed service setup included:
- Ongoing QA procedures and validation workflows
- Iterative batch processing and corrections
- Regular communication with the client team
- Continuous alignment on evolving annotation rules
- Scalable support for new dataset cycles
Outcome
Over time, the improved annotation consistency and QA structure helped the client streamline internal model development and reduce dataset preparation bottlenecks.
The cooperation continues to support ongoing experimentation and expansion into object detection-based workflows.
The client benefited from:
- Consistent annotation quality across 17,700+ images
- Reliable dataset preparation for multiple computer vision models
- Faster iteration cycles for model training and validation
- Scalable annotation capacity aligned with project growth
Few Words About Cooperation with Us
Working with highly variable real estate data required strong consistency and attention to detail. Mobilunity-BPO helped us structure our annotation pipeline and maintain quality across very different dataset types, which significantly improved our model iteration process.
