Home Virto Commerce blog Artificial Intelligence in Real Estate — Image Recognition Tools vs Toilet Bowls

Artificial Intelligence in Real Estate — Image Recognition Tools vs Toilet Bowls

Jul 10, 2020 • 5 min


A lot has been said and written about the importance of photos for real estate sales. Photos are the crucial criteria that influence the Ads Click-Through Rate (CTR) and the further conversion into a request.

Jakob Nielsen, a well-known expert in the field of website usability, writes in his book "Eye-tracking Web Usability" that "images are the most effective way to present information on a website, because people react to them instantly".

"Proper" (taken by a wide-angle, bright, juicy, without extraneous items) photos significantly enhance the chances of a quick and successful sale of the property. According to various data sources, the conversion "view - request - visit" increases from 30 to 70% in case of using of well-captured photos.

By the way, images too well shot and therefore not quite reflecting the truth are not a good option. People are often disappointed after what they see - a typical meme " expectation vs. reality".

If correctly done photos increase the probability of a quick sale, then poorly done ones will reduce these chances at least by half. It is complicated for a potential buyer to ignore the "bad" picture and believe that it is a worthy option.

There are even entire websites and SMM communities dedicated to terrible real estate images, for example, https://www.instagram.com/theterriblerealestatephotos

Backyard pool

Low-quality image of real estate. Beautiful backyard with a pool, really?

There is a deep correlation between the cost of the property and the quality of images — the higher the budget, the more significant the impact of the photo quality. Ideally - when both the photo and the description of the property correspond to each other (the picture has everything that is specified in the description). 

In the era of the COVID-19 pandemic, the demand for remote viewing of homes and getting information online has grown enormously. The value of visual tools, such as high-quality 3D/ HD tours, live online video tours, is increasing rapidly. Nevertheless, images remain a core and significant part of successful sales.

The importance of visual content

The owners of any successful portal or real estate agency know and care about the importance of visual content.

By the way, images of Scandinavian interiors look particularly attractive. If you wish to get an aesthetical experience - visit any real estate site in Sweden. Many of the pictures are so well made that they deserve to be published in glossy magazines (and it's not just about the Swedes' ability to work with interior design. If you look at how Swedish agencies photograph foreign real estate, the result will be the same — they focus on the visual approach).

Examples of Real Estate Images made by Sweden Agencies (qlstockhol, fastighetsbyran)
Examples of Real Estate Images made by Sweden Agencies (qlstockhol, fastighetsbyran)

Working with images is a distinct process, that is carried out daily by employees of our content department and takes up to 40% of their working time.

Out property catalog is generated from different sources, including:

  1. Properties for sale directly from the owners (FSBO).
  2. Properties for sale through partnership (by other agencies).
  3. Projects by developers (new buildings).

Therefore, the content editors should not only pay attention to FSBO photos, but also to check the images received from partners (which is more than 60% of the catalog).

The daily images workflow includes:

  • Discarding inappropriate photos (lousy quality, unacceptable content, illegal content — agent phones, logos, etc.
  • Selecting a photo for the property cover.
  • Order of images to be displayed.
  • Creation of captures and <IMG ALT> tags for photos (for SEO purposes).

If we do not do this, the main (cover) image of the property can be random (since not every company pays attention to the order of displaying, or simply does not transmit this data to us). So, we can receive unexpected results like photos with horses, toilets, ruins, or just poorly taken ones.

These are some examples of the photos that we deal with on a daily basis.

Also, we syndicate the catalog and distribute our content to various online and offline sources:

  • to other partners' directories
  • to property portals
  • to the MLS system (Real Estate Property Listings)
  • to promotional feeds (e.g., Facebook and Google remarketing)
  • for showcase printing 

That is why the risk of errors is quite high — a bad photo in the basis directory is automatically transmitted to other sources.

AI image recognition providers and services

To improve the publishing process and data quality, we decided to apply AI image-based technologies.

Nowadays there is a variety of Image recognition providers on the market, such as

  • AWS Amazon 
  • Google Cloud Vision AI
  • IBM Visual Recognition 
  • Microsoft Azure Computer Vision AI 
  • Clarify

Several ready-made solutions cover most of the options we require.  Some of them are already customized to the real estate sector. 

  • https://inten.to
  • https://restb.ai
  • https://www.sentisight.ai
  • https://vize.ai
  • https://imagga.com

Preparing images for a training data set

We have split the implementation into several stages (in order of priority).

The first and most important was the recognition and automatic tagging of the photo with the highest accuracy. We intended to use the information obtained through tags in this way:

  • Make a list of inappropriate tags. If an image is tagged by one of them, it cannot be a cover, or even added to the property slideshow. 
  • Create a list of priority tags for choosing a photo as the main one (cover).
  • Set up an image slideshow order based on tag information (e.g., if an image is marked as a bathroom, it should go after a kitchen or living room). 
  • Automatically apply features based on tags. 
  • Prepare a self- generated description (as a template-basis for detailed description) 
  • Discard unacceptable content (blurred photos, 3rd party logos, phones, advertising etc.)

The first step was to collect the proper number of photos to train our model. According to the service provider’s recommendations, it was necessary to collect as many images as possible to get the most accurate result.

For one tag, we needed at least 20 photos (in general, that was no problem for us).  Also, it was not necessary to strive for the best quality photos - the sources should be as close as possible to what we work with every day.

The next step was to prepare data for images self-tagging. We used one of the pre-trained services, so we did not have to do the primary manual tagging.  After uploading photos and processing them, we got the tag lists for revision.

We received about 20 tags of different types per each image. The most accurate data (probability from 100 to 80% match) included 2-4 tags per image.

We classified the obtained tags by types – primary (property type - living room, bedroom, yard, bathroom; condition - excellent, good, bad, terrible) and secondary (style, materials, interior items, etc.).

Another task was to identify rendered images. Since in our database about 30 % of the catalog are new buildings, it has always been a challenge.

It is assumed that rendered photos evoke less confidence. If the property has just come for sale, except for rendering and visualization, the developer has nothing more to show.  But if there is a show house and live property images are available – it is always better to use them).

During initial processing, the service could only determine 32% of renderings, so it had to involve human resources and tag more than 3000 photos manually. After the secondary processing, the accuracy increased to 67% (in the final test we managed to achieve up to 78%.  Now, if a photo is marked as rendering (and this item has live photos), we do not put it on the cover, but at the end of the slideshow.

We also managed to detect an image as a layout, plan, screenshots of the location, or even the price list.

Using trained model

Using the obtained information, we determined rules for main photo auto-detection and the order for creating a slideshow. Besides, we used these data to auto-name files and generate alternative descriptions of photos (to improve SEO and UX).

Our main goal was to cut off inadmissible photos for the cover, and the AI-tool performed well on the whole. We adjusted the algorithm several times.  The best outcome was shown in the case of strict rules. (e.g., if a photo is marked as unacceptable with at least 60% probability - we don't allow it to use it as a cover photo).

After about 2 months, the share of bad images used for covers decreased to 4% (before the adoption of the tool) it was 34% — every third object had to be modified).  So, it became possible to shift this issue into a post-moderation mode, and our content department began to spend 10 times less on it (about 4 working hours per month compared to 40 hours before).

We planned to use tag data to prevent the publication of unsuitable photos and select the best photos for the cover. Based on our ad clickability data, we made some assumptions about which photos would yield the maximum number of clicks. For example, we knew for sure that a photo of a swimming pool, taken "from the water" always attracts more attention. 

For example, the open conversion rate when using the first photo was 1.6 times higher than when using the second photo and almost 3 times higher than when using the 3rd one (without the swimming-pool).

Selecting the best photos for the cover turned out to be a much more complicated issue than filtering out unacceptable ones. For example, in the case of a pool is not enough just the rules of autodetection (if the tag contains a pool - take this photo), as one object may include several photos of the complex, showing the pool.

Based on our experience (and we are beginners, and tested this system for about 6 weeks), we can say that the more pool ones see on an image, the better this is for clickability. The situation is the same with the sea view — the more sea in the picture, the better (at the same time, the property itself must be visible).

At the same time, an item with a sea view in the main photo is clicked more often than the one with a swimming pool in the main photo (sea always hits the pool).

Another "clickbait" feature is a beautiful green area, garden or landscaping. If an offer has neither a sea view nor a pool, the third most valuable option is to use pictures with green spaces.

In general, the challenge of selecting the best photo for the cover is not yet complete. We are still testing different sets of tags and their sequences for the best outcome: the more "valuable" parameters, the more difficult it is to pick the best image. We also rotate "good covers" — we change one decent photo to another every 2 weeks. (By the way, this simple trick is used by some companies placing their ads on property portals. To boost interest and exclude an item from the "blind spot" zone (when a user has seen this image many times, he is not interested and continues to ignore the ad), companies manually change the cover photo from time to time.)

Enrich and sell — AI data enrichment experience

Our team pays much attention to SEO traffic. For this reason, we created quite a large number of landing pages for different types of queries to obtain SEO-traffic. Among them are many low-frequency (or so-called long-tail) keywords, such as location + property type + property features.

Luxury villas for sale in Finestrat with a seaview and pool

We also utilize many filter options for advanced search and enhanced customer experience.

There is always a problem that not all items are labelled with the most significant number of tags they could have. Some of the features can be used automatically to avoid being dependent on the "forgetfulness" of the content editor. But you still need to set most of the tags manually. Processing items from the 3rd party directory also requires you to apply manual tagging to prevent valuable data loss.

As a result - in case of deep search (if the user applies more than 4 criteria in filters), there is always a chance that the final output is not full. And the more in-depth the query, the higher the conversion rate can be (as we deal with a high level of client involvement).

To improve the output, we also used AI image-based solutions. To create a large number of different landing pages, we use the Virto Commerce catalog module. This is a pretty flexible solution, making it possible to integrate digital data of any type and size. This was extremely important for us because we engaged a third-party service to expand the main property catalog.

As we mentioned earlier, after running the photos through the database, we got up to 20 tags per image. At the same time, the number of frequently repeated tags was not that high — about 100. 

We've added tags and features that may be useful to both the client and search engines from this list. 

We have combined some of the tags into clusters to use them as a group of synonyms.

We started the process of automatic tagging based and got a good result —   doubling the number of items in the output (which means at least 50% of useful information we have not used before).

After that, we started the process of automatic tagging based on photo data and got a good result - doubling the number of objects in the output (which means at least 50% of useful information we have not used before).

Another key improvement in content publishing was the creation of self-written texts to speed up the work of the copywriter. Preparing text for the description is the most time-consuming process, which can take up to 50% of the time. The editor needs to go into all the details of the property, look at all the photos and mark all the important points.  We decided to speed this up by providing content editors with a list of tags found on photos to create more accurate and vivid descriptions.

After uploading a photo, you can see the list of found tags that can be applied when writing text.

Example list of tags:

  • Villa 
  • House 
  • Spacious 
  • Garden 
  • Lawn 
  • Mountain View 
  • Living room
  • Modern design
  • Contemporary design 
  • Natural light 
  • Pool 
  • Outdoor amenities

And the result

We recently started testing this system. We're still fixing its faults (for example, sometimes a list of tags and features is too long, and the content editor has to delve into it and remove unnecessary information). And of course, it does not replace manual work. But descriptions became more saturated with details, and they are produced faster.


We can consider our experience with AI image-based services as satisfactory. 

Photo is a valuable source of information for real estate companies, and one of the main drivers influencing the client's choice. We certainly recommend real estate companies to adopt AI image-based services and use them. They can be especially useful for web-portals dealing with UGC content (to reduce the amount of unacceptable content and improve the user experience when creating ads).

Of course, the implementation and use of solutions based on AI image requires a solid preparations and funding and support in the future. But the result you can achieve will definitely leave all your competitors behind. In the highly competitive real estate market, in order to win, you should provide a high-level user experience. And AI will definitely fight for the attention and loyalty of your clients.

About the project

VirtoProperty is a highly experienced and dedicated real estate company that combines local market competence with international contacts. Besides, VirtoProperty can harmonize old-fashioned business values and ethics with the latest state-of-the-art technology tools, as well as a profound comprehension of modern markets, consumer behavior, and internet-based marketing techniques. The company is headquartered in the most demanded resort area of Spain — Alicante, Costa Blanca.


Schedule a quick demo

You might also like...
Multi-Store eCommerce: The Ultimate Guide & Top Platforms in 2024 Multi-Store eCommerce: The Ultimate Guide & Top Platforms in 2024
 Mary Gabrielyan
Mary Gabrielyan
Mar 1, 2024 • 10 min
The Essentials of PunchOut Catalogs in B2B Ecommerce The Essentials of PunchOut Catalogs in B2B Ecommerce
 Oleg Zhuk
Oleg Zhuk
Feb 22, 2024 • 15 min
Understanding Serverless eCommerce Platforms: Complete Overview Understanding Serverless eCommerce Platforms: Complete Overview
 Oleg Zhuk
Oleg Zhuk
Jan 14, 2024 • 15 min
Copyright © 2024. All rights reserved.