Overview:
Homster is a cloud-based Artificial Intelligence service that empowers home sellers to enhance property sales by generating personalized virtual tours from floor plans. These virtual tours are tailored to the preferences of potential homebuyers, resulting in faster sales at higher prices. To achieve this, Homster collaborated with Co-one, a data annotation service provider. Our objective was to accurately annotate floor plans based on over 15 object names, such as balconies, kitchens, and living rooms, to improve the performance of Homster's AI service.
The Problem:
Homster faced a significant challenge in ensuring that its AI service delivered accurate and personalized virtual tours. This required precise labeling of floor plans according to various object classes. With more than 15 object names to consider, ensuring consistency and accuracy in annotation was paramount. The potential for inaccuracies could impact the quality of the virtual tours and, consequently, the success of property sales.
The Solution:
Co-one provided a comprehensive solution to their data annotation needs, focused on enhancing the accuracy of Homster's AI service:
Co-one commenced the project by developing a detailed guideline, meticulously outlining the rules and standards for annotating floor plans. This guideline, created in close collaboration with Homster, aimed to address edge cases and questions, ensuring alignment on annotation rules. This rulebook clarified the meaning and guidelines for each of the more than 15 object names, ensuring a consistent and standardized approach to annotation. Customer feedback from Homster played a pivotal role in shaping these guidelines.
Homster transferred their data to Co-one, who subsequently leveraged a crowdsourced team consisting of over 1,000 annotators. Each annotator was assigned a specific object to label on the floor plans, streamlining the annotation process and ensuring a rapid turnaround.
To further ensure accuracy, a cross-validation system was implemented. Errors detected within this system prompted Co-one's expert reviewers to verify the annotations, ultimately guaranteeing their correctness.
Upon the completion of the annotation process, the annotated data was meticulously prepared for delivery to Homster.
Our operations team undertook an in-depth analysis of the annotated data, providing Homster with reports tailored to the specific requirements.
We remained committed to achieving a minimum accuracy rate of 90% in labeling. In cases where accuracy falls below this threshold, all data undergoes a rigorous review process.
Homster also opted for "Simple Reporting, and Expert Control" services from Co-one. This service involved data controllers with expertise in the field meticulously checking the tagged data. The project report included object and annotation distribution graphics, as well as charts illustrating the distribution of images by annotated object types and numbers.
The Result:
The collaboration between Homster and Co-one yielded remarkable results. Homster experienced a substantial increase in the accuracy of its artificial intelligence service. The precision with which floor plans were annotated will enable the generation of personalized virtual tours that precisely match the preferences of potential homebuyers. Demonstrating the power of AI combined with effective data annotation in the real estate market.
Customer Review:
Our existing deep learning model needed to be retrained with new data that the model had never seen before. The accuracy of the labels was the most important factor for us, and thanks to the strong communication of the Co-one team, we got our labels without any problems. On the other hand, being able to handle a process that would take 2 weeks in 1-2 days accelerated our work a lot. We can now get results with an accuracy rate of over 90% in examples where we could not get results before the training.
Selim Ceylan, Computer Vision Engineer, Homster