Hacker News new | comments | show | ask | jobs | submit login
StreetScore scores a street view based on how safe it looks to a human (mit.edu)
120 points by lbotos on July 25, 2014 | hide | past | web | favorite | 42 comments



What I would like is something that took street photos and safety info and somehow deduced what visible characteristics correlated with safety.

A while back I was reading a mailing list thread where someone claimed that adding more street lighting didn't improve safety. This struck me as bullshit.

I did some Googling to dig up some fact to show just how right I was. Apparently, not so right at all. There were assorted studies that showed (or suggested, not sure) that adding more lights might actually make a street more dangerous. They can make it easier for a hidden thief to case a place or a target, and some lights create enough glare that it becomes harder to see past a certain distance.

It got me thinking that popular notions of how to judge safety may be very wrong, and that reality may be quite counter-intuitive.

Another example: I grew up in a city, lived in apartment buildings. Anyone who had any money moved out of an apartment and moved into a house. Hence, anyplace that was all houses was a good neighborhood.

Years later I discovered how totally wrong this was and marveled at how such impressions get formed.

Having better evidence about what factors make a location safer, and to what extent one can accurately determine safety by visual inspection, would be very beneficial.


I'm using a throwaway account so that I don't inadvertently reveal my location.

The area where I live is on the map. I found it very interesting that the safety ratings seemed to be closely correlated with lighting. Street View took pictures of one popular outdoor area near me on a rainy day. Of course, it was deserted due to the weather, so the entire area received safety ratings of around 0. However, on a sunny day, the area is filled with families, and I'm sure that it would have consistently received a 10.

Likewise, my building received low ratings when compared to the surrounding area due to the presence of an alley (which is actually a gated entrance). However, no crimes have taken place there for a very long time. An area just up the street received a rating close to a 10, yet armed carjackings have taken place there before.

At least from my own (rather anecdotal) experience, it's fascinating that the factors which cause people to perceive something as safe or unsafe can actually be far from reality. Anyway, StreetScore is an Awesome piece of research from the ML. I'm interested in seeing the direction that it's headed.


I'm confident that incorporating pictures of the same place at different times/days (like Google Map's wayback machine) would boost the accuracy, but I'm not too sure how you would need to change the model.


We have odd perceptions of safety. If you look at the website, you'll see the first photo labelled 0.51 for safety. But, honestly, what's the worst that could happen there? There'll be no theft, because there are no houses. There will be less street crime, because there are no people. In fact, the one rated highest is probably the least safe, as it would likely be targeted by thieves looking for high valued property in houses from a nice neighbourhood. As far as murders and violence is concerned, these crimes are mostly targeted and it doesn't really depend on where you live.


You seem to discuss much different aspects of safety - I'd assume that the article refers to the threat to a potential passer-by, not of the general crime rate in that area, which would include thefts from property.

Also, safety is not neccessarily correlated with a crimerate - if place A has 100 times lower rate of street crime than place B, "because there are no people" as it has 1000 times less foot traffic, then place A is 10 times more dangerous to a person who actually happens to be walking there.


That's really interesting. I remember when this [1] article came out, showing the difference in streets' income level with the greenery in the street, I found it amazing. Once I started thinking about it, I figured it should have been rather obvious.

So, the way to do would be to first figure if there's more crime in lower-income streets (likely. Right? No data to support me here) Then using the 'foliage detection algorithm', which you can use on LIDAR data available from USGS, which urban neighborhoods have less, and more crime. Do some tests, to see if your hypothesis holds and water at all, and it should be a solid test.

http://mashable.com/2012/07/18/google-maps-income-inequality...


I would assume foliage is correlated to wealth. The most affluent areas of town in my city are generally placed on half acre lots (the old city has some that are on acre plus lots inside the city), upper middle class are on the 1/8th to 1/4 acres lots. I live in the original limits of the city, my house is 100 years old and sits on 1/16th of an acre. I'm solidly middle class and my neighbourhood feels it. The biggest thing to happen in my area in the past few months was a car accident on the main road just today. The last was a domestic in a 3-unit rental (IE cheap rent).

The issue with this theory is that my city has a strong industrial sector. So you can actually get great big 1/4 acre housing lots for really cheap. You're sandwiched between train tracks and an industrial machine shop. So be prepared for lots of crack deals on your streets (not even joking).

Again, going purely by foliage, my friend had a fantastic place in the city. 1/8th acre lot, amazing location in town. Hookers on the lawn. Ever had to yell at a pimp that you called 911 with a 3 month baby in your arms? She had a house worth half a mil.

The biggest issue with "perceiving" danger is that it's not real danger. Sure, there's a lot of dangerous looking parts of town. I know, I work doing home renovations. I've been there.

I worked in the roughest area of town. We had expensive tools, we had scrap aluminum (I had two 4ft pieces weighing 30lbs a piece and were easily carryable. Scrapped for $70). My boss was freaked, worried something was going to get stolen. Except, we were completely safe. I'd say we were on the safest street in the damn area. Why? Because all the hookers going home in the morning were scared shitless of one house. That house had a black guy who puts NFL line backers to shame sitting on the porch all day, every day as people came and went. When a homeless guy had a freak out at the end of the street, he's on his way. Why? Because a guy who's likely dealing drugs or worse, and presumably doing pretty good at it to have a bouncer, doesn't want shit happening on his street.

The Hells Angels club is in one of the poorest areas of town. When MS13 started showing up in town in the poor areas, their area ended up with a bunch of no-suspect assaults on gang members and it's been a couple years since I saw anyone who looked like they would belong to MS13. It would be a really bad area to live if you're Hispanic, but if you're white. It's probably one of the safest areas in town for you.

We as humans are pretty shit at evaluating actual risks. Our fears play into it too much, and our greatest fear is the fear of the unknown. I grew up in a very rough area of England and I moved to Canada. The "roughest area" has nothing compared to middle-class suburbia where I lived. At 2am in the morning am I afraid I'm going to get the shit kicked out of me by drunk hooligans? No. At 6pm on a sunday am I afraid I'm going to get mugged at knife point by a bunch of 15 year olds? No. People on my street feel "in danger" because there's teenagers playing on skateboards.


> I would assume foliage is correlated to wealth

Likely. Lawns have pretty much always been conspicuous consumption: Versailles is sometimes considered the source of the first modern lawn, as opposed to a pasture.


In general, but the correlation isn't perfect. E.g. compare Toledo, Ohio to Toledo, Spain. http://www.newworldeconomics.com/archives/2010/022110.html

Some places will have much more green area just because they have much more open space, or empty areas. A more compact or older city will have less space, less green area. But not necessarily because poverty.


How safe does this street "look" to a Human?

NOT

How safe is this street?

Very important distinction - I would suggest changing the name of the project to "PerceivedStreetScore".

What is now needed is actually part 2 of this project - one that pull records from City/County/Federal crime stats, overlays them with a Map and then actually produces a "ActualCrimeScore" for each of these highlighted streets.

Then compare the two scores and write an (excellent) definitive paper titled "Factors that influence Human perceptions of safety - Lessons for City Planners"


If you look at the 2013 paper from this same group you will find that this is exactly what they did (http://pulse.media.mit.edu/papers/).

Even further, they used geospatial statistics techniques to control for the effects of the income, area, and population in an area, and hence, developed a model that incorporated not only perception, but also demographics, when predicting the location of homicides.


From the site's FAQ:

>StreetScore is a machine learning algorithm that predicts how safe the image of a street looks to a human observer. We trained StreetScore to predict perceived safety using a 'training dataset' consisting of 3,000 street views from New York and Boston and their rankings for perceived safety obtained from Place Pulse — a crowdsourced survey.


Don't forget to factor in time as well, it would be accurate to judge a high crime street as safe during the day if all of the crime happened during the night.


Not to mention that one street may very well be safer for one individual compared to another depending on any number of variant factors - race, gender, build, etc.


I was expecting a traffic fatality map, something like this: http://riskyroads.org/


Corrected headline: "How safe would humans say this street looks?"

Not the same question at all.

While the research into computer vision and machine learning is interesting, as a human I'd like to know when my intuition about safety is wrong, not just have a machine that can mimic my intuition.


Unfortunately, once you place the imprimatur of machine learning upon it, people will tend to conflate "perceived safety" with "actual safety." At the extreme, it becomes a way to create an apparent but false confirmation of the prejudices of the crowd.

This is one of those rare cases in computer science research where there is an actual ethical component to the work being done; while machine scoring landscapes according to their relative attractiveness is fairly harmless, scoring them according to "perceived safety" without mentioning the obvious limitations thereof is heading towards trouble. As I recall, there was a crowdsourced app last year that supposedly would inform its users if they were in "bad" neighborhoods (using the opinions of other users to tag areas), and it received a wave of criticism for essentially similar reasons.


If it was crowdsourced, why would you need an app to know if you were in a bad neighborhood? Your perception should be roughly similar to those who had contributed their opinions as crowd source data. Kind of like needing an app to tell you it's raining when you're standing in a downpour.


The act of using the app can increase that perception though; if you get mugged whilst checking your smartphone in the middle of the street it's a pretty good indicator the area isn't safe :-)


"Would humans say" is also subtly different from "do humans actually think". The Williamsburg bridge in NY has a particularly odd pattern where pictures showing steelwork have particularly "safe" ratings and pictures showing river panoramas have extremely low ratings. Maybe HNers other than me have an idea of why that might be, but I find it hard to believe the steelwork is incredibly reassuring and the view of Manhattan behind utterly terrifying as the polarized figures would have you believe


If you can automatically get data "how safe it looks" for all addresses, then it's a way to actually measure the [anti-]correlations with actual crime happening there.

We had the actual crime data already, and now we could compare it with the perception to try and identify the reasons for mismatches. Not that identifying reasons is simple, though - with machine learning it's far simpler to get a system that makes good decisions than to get an understanding on why it makes those particular decisions.


Would be fascinating to see how these data correlate with crime statistics, in particular for violent crimes. Perhaps more fascinating would be if changes in these scores could actually predict changes in crime levels. I'm sure the authors–or their colleagues-are already working on this.


This isn't the best visualization, but I just spent a couple minutes overlaying the New York map with Trulia's crime risk heat map[1,2]. I think the important takeaway is that the green dots visible in the red areas can be interpreted as false negatives for danger. Of course, this is a really, really rough sketch so take it with a grain of salt.

[1]: http://www.trulia.com/local/new-york-ny/tiles:1|points:1_cri...

[2]: http://imgur.com/HOuJFKZ


The crime map seems awfully suspicious to me. It appears to be more of a population density map than anything. All the red areas south of 34th St. just look like common places for people to hang out (and drink alcohol).

I think what people really want is a crime map that plots crimes committed at random, rather than drinking buddies beating each other up.

I also took a look at vermont, and the red areas were because of things like "traffic stop" or "e911 hangup".


Check this map instead http://i.imgur.com/QwvATt1.jpg The authors already did that analysis, but they incorporated population, area, age, income and perception of safety into a statistical model used to predict violent crimes.

It is important to go beyond the bivariate case here. Comparing streetscore to crime is too simple to be meaningful if other co-variates are not taken into account. For the full description (http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjourna...)


It's actually well understood which landscape features contribute to or prevent crime. Things like sidewalks and bright street lamps have a positive effect (no surprise there). Bushes, certain fences and other objects that obstruct the view from the road can have a negative effect. As a reporter I once did a ride-along with a police sergeant who took it upon himself to cut a set of hedges (with the property owner's permission, of course) that helped facilitate drug deals and prostitution.

The point is, I wonder if the machine learning approach used here is overly complex. After all, the set of environmental factors affecting crime is so well understood and thoroughly researched that you could focus on detecting tried-and-true things such as sidewalks. This would entail applying a clear set of rules instead of using the relatively unsupervised approach with training data. To be fair, ML is a complicated subject and I'm not an expert; maybe their approach draws heavily on these things.

EDIT: I understand that perception, rather than the actual crime rate, is the focus of this research. Still, there seems to be a tight correlation between the features that are known to be dangerous and those that appear sketchy. The major ones - an absence of lights, few walkways, etc. - are obvious to most pedestrians.


Things like sidewalks and bright street lamps have a positive effect (no surprise there).

There may be some surprise there.

The correlation between street lighting and safety is not obvious.

"In 2008, PG&E Corp., the San Francisco-based energy company, reviewed the research and found 'either that there is no link between lighting and crime, or that any link is too subtle or complex to have been evident in the data.'"

http://www.bloombergview.com/articles/2013-02-24/turn-down-t...

http://www.popcenter.org/responses/street_lighting/2


One thing they should consider doing is using the average of the surrounding points to rise or lower the score of a particular spot. For instance, near where I live there a number of examples where there is a green dot on one side of the street and a red on the other, just down the block they reverse, this doesn't make any objective sense at all.


I'd be cautious about that. You can easily have a dark foreboding alleyway just off a bright, well-lit street. Safety gradients (perceived and actual) can be quite steep.


Interesting program. This info needs to be taken with a grain of salt, or somehow correlate to existing statistical information on certain neighborhoods - I doubt ~400 W 63rd Street in Chicago is a 9.1. The theory of it is great for visualization, and I'd love to see more incorporated data in the future.


I remember clicking through those same pictures a few days back and I was under the impression that I was telling the observers what pictures I thought were more beautiful, not safer.

Did they just re-label the same data here or is this another separate program?


Yup. The data is from here http://streetscore.media.mit.edu/data.html and it's used in all PlacePulse surveys. http://pulse.media.mit.edu/ You can choose which question you want from the big menu at the top.


I guess this would be a good data point in building a "safety" score, but by no means would I consider this telling me if an area was "safe."

Crime statistics, thoroughfare, property value trends, population density and other environmental factors are just a few of the things I would consider looking at as well if I was compiling data.

Interesting start to an interesting question. I'm just not comfortable with this "judging a book by its cover" mentality to this...


The police in the UK actually have a map of all crimes that have taken place recently in a neighbourhood [1], I think this is a better way of going about the issue, crime rates seem to correlate more closely to safety than how a place looks, I think.

[1] http://www.police.uk/


I think you fail to understand the point. The point is not too replace a map of crime, but to help develop variables that can help explain the location of crime. Look at what they did in their 2013 paper (http://pulse.media.mit.edu/papers/) that is exactly that.

The problem they are trying to solve is that there are no good measures of perception to evaluate whether the environment affects crime or not. This is an old question for which there has been much research, as it is explained in the point 7 of their FAQ. http://streetscore.media.mit.edu/faq.html


Being from Detroit, I found the map particularly interesting to look at.

Personally not a fan of the UI currently. However there could be a lot done with the itself- such as providing safer routes by utilizing the data.


How badly does gentrification break this system? After all, any cred you have as a trust fund hipster is lost if your home looks like it's in a "safe" part of town ;)


It's a good start. With individual pictures it's sketchy, but on the aggregate the map I looked at trends with my own perception of the area, generally speaking.



You should also checkout project Pantheon: http://pantheon.media.mit.edu


They should add an urban planner to the team.


It seems to classify open areas as unsafe and photos with lots of variation as safe.




Applications are open for YC Winter 2019

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact

Search: