That’s a cute house!
Teaching a machine to judgmentally assess “Curb Appeal”
I love looking at houses
I have an obsessive habit of checking Redfin listings and judging strangers’ houses. It’s not very complex analysis. Usually, it’s just like “Oh that’s a cute house!” or something along those lines. But after years of this, I think I have a pretty clear idea of what kind of houses I like.
Lately, I’ve been wishing the Redfin interface gave me a more unique and specific set of filters, more than price, bed, bath, etc. One I’ve been wishing for forever was filtering by Walkscore, which seems completely doable since they already own Walkscore anyways.
The other big one is an aesthetic-based search. At the very least it would be nice to be able to search by house style (bungalow, cottage, etc) especially since this data is already available in the listing usually. But even better, it would be nice to say “only show me the cute houses.”
Obviously, MLS does not include a cuteness rating for each listing since it’s quite subjective. And it would be a pain to manually tag them. But most listings have a photo, and computers are getting pretty good at labeling photos now.
Teachable Machine
Google’s Teachable Machine lets you “train a computer to recognize your own images, sounds, & poses.” The no-code interface is super friendly and easy to play with (If you’re interested in more details, this tutorial is a pretty good walkthrough). I have been wanting to try it out for a while but didn’t have any good reason to until now.
To build my model, I grabbed a bunch of random photos from around Los Angeles and sorted them into “cute” or “not cute” folders then uploaded them into Teachable Machine then trained the model. I made sure to sample from all over the city so it wouldn’t just like label all the expensive west side houses as cute. (Actually, there’s a lot of really ugly expensive houses on the west side too). Eventually, I also added a “not a house” class as well because when I was testing it with the webcam it kept telling me I was “not cute” and that made me a little sad.
After building the model, it lets you export it to be used elsewhere. And in this case, Shawn hacked together this little web app using the model.
I was pleasantly surprised by how well it actually works as a prototype. Of course, it’s a very small sample. And like all models, very biased. It’s based on what I personally think makes a house cute, so you might not agree with it.
In this little example, it’s judging photos one at a time, but I imagine it wouldn’t be that hard to go through a whole set of listings and then filter them accordingly based on the model. So anyways, Redfin Engineering, maybe you can take it from here?
