“No matter which annotation service provider we use, the result would be the same”.
That’s what we thought initially.
Then, ABEJA brought us a new insight as the work went along.

About LIFULL

LIFULL Co., Ltd., aspires to become the best life-event database and solution company in the world. They operate a real estate and housing information aggregation web service “LIFULL HOMES”. As Japan’s top (*) provider in terms of the number of available properties, their services cover a wide range of information including renting, buying, building, renovating and real estate investments. * Surveyed by Sankei advertising agency (January 7, 2019)

LIFULL pro-actively develops new machine learning based services leveraging their data assets including property specifications, images and floor plans.

Interviewee

(Right) Mr Takashi Yamada, Director
(Left) Mr Satoshi Shiibashi, Data Scientist

The background

Leveraging LIFULL’s property data to “visualize” real estate information

“We are developing a new system combining Machine Learning technology and our data such as property specifications, images and floor plans” says Mr Yamada, Director of LIFULL.

It’s a part of “visualization” of real estate information, something LIFULL has always been pushing forward since they established the company. Developing a new system has been one of their long-time challenges.

“There is very little information available for house purchasers despite the fact that house prices are several times more than people’s annual income” remarks Mr Yamada. They believe LIFULL’s mission is to solve this “information asymmetry”. That is why they focus on adopting technology that delivers information about properties, prices, quality and real estate agents to their service users.

For example, LIFULL provides a service called “Hold up high and search”. Using AR(Augmented reality) technology, you hold your smartphone up towards a housing complex you are interested in, and it tells you the current rental or sales information through your camera screen. They have also developed a service called “PRICE MAP”. It uses Machine Learning technology to estimate and visualize the property prices.

“We have created these new services, but still there is room for improvement” points out Mr Yamada. One of the issues is regarding the process when a customer makes the final decision on a property. Even though people are able to gather a lot of information online, they still need to go through the process of actually visiting the place to view houses and go to real estate agents to sign the contract. It’s a conventional processes yet it still resides.

To make this process easier, LIFULL focuses on real estate data such as property specifications, housing images and layouts. These were all data assets collected through their work trying to “visualize” housing information. It has a strong advertising impact that people are now able to find this real estate information on most of the 7 million properties introduced on the LIFULL HOME’S website. “We began to wonder if we could use this information to develop a new service leveraging Machine Learning and image recognition” explains Mr Yamada.

The solution

In search of a reliable partner who can create high quality training data

“To create high-precision AI models, the process to create high quality training dataset is also very important” explains Mr Shiibashi, Data Scientist of LIFULL. This process is called “Annotation” and it is about assigning labels on images to let it learn what they are.

This is done manually by human editors and is extremely resource intensive. LIFULL did not have enough resources inhouse, so they decided to outsource this data annotation process.

When LIFULL was looking into annotation service providers, they thought “the quality would be the same anywhere, since the process of checking the images and judging are done by human workers.”

Moreover, it needed to be a reliable service provider. LIFULL felt that ABEJA could understand their intentions well because of its years of experience in developing Machine Learning models. During the data annotation process, annotators are given an instruction manual, but in practice they often come across cases where it is hard to judge. If annotators make a wrong judgement at that point, it could cause a negative impact on the model training.

“ABEJA has a lot of knowledge and experience on Machine Learning so we expected the annotation quality to be very high” says Mr Shiibashi.

LIFULL decided to use “ABEJA Platform Annotation”. It is one of the functions of “ABEJA Platform”, a product that covers end-to-end processes of Machine Learning. In addition, they decided to use another agent that provides cloud-based data annotation service.

Two companies lined up, then ABEJA’s solution showed supremacy

“To tell you the truth, we intentionally mixed some dummy data that had already been annotated” says Mr Shiibashi. LIFULL wanted to check the quality of annotation between these two companies. At this point, both ABEJA and the other company delivered correct outputs and there were no issues.

However, as LIFULL went along with the annotation process with ABEJA, they realized they could do things very efficiently using the platform. Generally, when you outsource data annotation, there are still some parts that need managing and supervising. For example, they need to prepare the base of the annotation process, or to prevent overlaps or incomplete annotation.

ABEJA’s platform already includes these and provide them as a whole service. Once they upload all the images from the platform to the cloud, LIFULL only needed to wait until it was completed. “Since ABEJA also has internal reviewers to check the annotated images, we didn’t need to manage the progress. This helped to reduce our workload” says Mr Shiibashi.

ABEJA’s solution brought extra advantages - saved cost and time

On top of that, LIFULL originally expected they would need 10,000 training data to achieve their target accuracy of their AI model. However, they attained sufficient training results by just 1,000 training data. “It saved our cost and time. It was all thanks to high-quality training data made by ABEJA’s diligent customer support” expresses Mr Shiibashi.

“It is definitely worth using a ready-made platform to get the job done quickly while keeping the workload minimum. At first, when we were selecting the vendor, we didn’t have much of a preference, but now we are really glad we chose ABEJA.”

The future

“A global platform”

Using Machine Learning technology, LIFULL keeps working on visualization of all real estate information even further. In the future, LIFULL aims to create “a global platform” which will enable users to look up world-wide real estate information. People will be able to view houses, sign contracts and settle payments all online and in their own languages. This will help them find a place to live, and it will also make international real estate investments much easier.

To create this “global platform”, it’s essential to further enrich information and functions so that people no longer need to physically visit the sites before they make decisions. “That is why we believe it is important to leverage Machine Learning and enrich information by our new services” expresses Mr Yamada.

Each country has its own real estate laws and regulations, and they expect to face many difficulties on the way. However, LIFULL plans to advance step by step to make their vision of “a global platform” come true.

Please feel free to ask us if you have any questions about our products.
We will get back to you promptly.

PAGE TOP