I spent some time over this week looking at various technical angles. I'm going to keep searching, but the conclusion so far is that any serious technical solutions to finding, sorting, and cleaning trash are the intellectual property of private corporations. I did find TrashNet, a collection of images and labels for a neural network database.
https://github.com/garythung/trashnet
I found TrashNet by reading an academic paper that made use of their library. Researchers in Turkey were able to take the library and apply a series of more complicated algorithms. They achieved around 95% accuracy with the best algorithm and expect that they may be able to achieve more in future work. What's very interesting to me, though, is that two students in California put out a free database that enabled researchers in Turkey to push the state-of-the-art just a little further. Now this is the kind of thing I'd like to see more of!
http://itnas.civil.duth.gr/inista2018/proceedings/inista2018_SML2.pdf
When you read an academic paper, you get their references and that can help you (with the aid of scholar.google.com) to track down even more great work. One paper mounted a camera on a street sweeper to collect images of trash "in the wild" and sorted/labeled objects with the help of software they wrote. In their summary of current work, they mention a number of projects that are making forays into automated trash identification and sorting while stating that they suffer from lack of data or of insufficient types/numbers of categories. This motivates their own work to create their own data set. But then they don't share their data! (I bet that a way to get the data could be found if someone were to e-mail them directly.)
https://arxiv.org/pdf/1710.11374.pdf
There are a number of articles that looked particularly interesting. One on an Android app that automatically labels trash in images taken from your smartphone looked particularly intriguing both because of the app and because of their new data set. However, the paper is locked behind a paywall and I don't have academic access anymore. (lame)
What I'm finding by tracing links and papers around, though, is that the effort of autonomous trash clean-up is very much in its infancy. Good work is being done, but data for analysis is scarce and getting more of it takes a lot of time and money. It seems mostly to be student projects or student labor. However, when the data does become available, researchers seem to be making good use of it to test their theories.
I also like that they're making a real effort to include the public. They've given the floating facilities personalities and you can download an Excel spreadsheet with quantity estimates of what they've collected in each bin.
The two things that would make these things even better is if they were able to catch things that didn't float at the surface and if they could discriminate. (I have to be careful because a: I'm certain they already have a long wish list of things they'd like to add given the funding and b: why knock a good thing.) At any rate, this is a simple device doing simple collection and so, while it's the perfect fit for its location and purpose, it's not a high-tech solution meant for the wild.
https://www.mrtrashwheel.com/
TrashNet
TrashNet was a Stanford student project from two years ago. Two students captured and labeled over 2,000 images of trash of various types. Their primary focus was on types of material (i.e. plastic, glass, cardboard). The students then used their database in an AI analysis and got around 75% accuracy. As part of their effort they created a GitHub account and posted most of the data to it. There are limits to how much data can be stored/transferred there so they host the raw database on Google Drive.https://github.com/garythung/trashnet
I found TrashNet by reading an academic paper that made use of their library. Researchers in Turkey were able to take the library and apply a series of more complicated algorithms. They achieved around 95% accuracy with the best algorithm and expect that they may be able to achieve more in future work. What's very interesting to me, though, is that two students in California put out a free database that enabled researchers in Turkey to push the state-of-the-art just a little further. Now this is the kind of thing I'd like to see more of!
http://itnas.civil.duth.gr/inista2018/proceedings/inista2018_SML2.pdf
When you read an academic paper, you get their references and that can help you (with the aid of scholar.google.com) to track down even more great work. One paper mounted a camera on a street sweeper to collect images of trash "in the wild" and sorted/labeled objects with the help of software they wrote. In their summary of current work, they mention a number of projects that are making forays into automated trash identification and sorting while stating that they suffer from lack of data or of insufficient types/numbers of categories. This motivates their own work to create their own data set. But then they don't share their data! (I bet that a way to get the data could be found if someone were to e-mail them directly.)
https://arxiv.org/pdf/1710.11374.pdf
There are a number of articles that looked particularly interesting. One on an Android app that automatically labels trash in images taken from your smartphone looked particularly intriguing both because of the app and because of their new data set. However, the paper is locked behind a paywall and I don't have academic access anymore. (lame)
What I'm finding by tracing links and papers around, though, is that the effort of autonomous trash clean-up is very much in its infancy. Good work is being done, but data for analysis is scarce and getting more of it takes a lot of time and money. It seems mostly to be student projects or student labor. However, when the data does become available, researchers seem to be making good use of it to test their theories.
Mr. Trash Wheel
There's a project I'd previously heard of that filters trash floating on the surface of a river in Baltimore. "Mr. Trash Wheel". The concept is pretty simple, skim trash at the surface into a conveyor belt that lifts it up into a bin. Either the river turns a big wheel to power the belt of solar panels pump water onto the wheel to turn it. Simple. Elegant. And effective! Although very slow, it's captured over a million pounds of "stuff", including 150 miles' worth of cigarette butts, that would have otherwise ended up in the ocean. They take everything they collect to an incinerator that burns it all up to generate electricity. I really have to tip my metaphorical hat to them as someone clearly had a vision and a team came together and got something done.I also like that they're making a real effort to include the public. They've given the floating facilities personalities and you can download an Excel spreadsheet with quantity estimates of what they've collected in each bin.
The two things that would make these things even better is if they were able to catch things that didn't float at the surface and if they could discriminate. (I have to be careful because a: I'm certain they already have a long wish list of things they'd like to add given the funding and b: why knock a good thing.) At any rate, this is a simple device doing simple collection and so, while it's the perfect fit for its location and purpose, it's not a high-tech solution meant for the wild.
https://www.mrtrashwheel.com/
Thoughts
Over the last few weeks, I've spent quite a bit of time looking at ocean clean-up sites. Most are appealing directly for money while showing pictures of volunteers on the beach cleaning and/or images of animals tangled in nets. I do believe they are serving an important role and I also believe that volunteers on a smattering of the world's beaches are probably not going to do much. Most are also advertising their focus on political solutions, which I think would probably be the biggest achievement in terms of impact on the future.
I have not found, though, much in the way of engineering other than "The Ocean Cleanup" and "Mr. Trash Wheel". The problem (and the oceans) are just too big and too expensive. It's a real problem since evidence suggests there could easily be millions of people who want to help, but there isn't sufficient funding to employ people and buy the hardware to get anything done.
I do think that one area I might be able to donate my time is to generate a "trash database" like TrashNet. I was happy to see how two people were able to influence people in Turkey and nudge things forward a little bit. Their database was a great start and maybe what I can do is to help nudge the database forward.
I'm thinking about some mixture of TrashNet and the "images of trash in the wild". For instance, I could walk around looking for trash. When I find it, take a picture of it where it lays. Then collect it and put it on a clean board and take a picture so it's isolated from the background. I could measure its dimensions and weight. Then I could label all the images by type of background (asphalt, grass, sand, etc.), type of material (paper, plastic, glass, etc. ), type of trash (cigarette, bottle, wrapper, etc.), and so on. It would take me a while, but if I spend every Friday (within reason), I might have a good database collected over time. Then maybe researchers somewhere out there would be able to apply their machine learning and robotics algorithms and maybe (maybe) someday a series of AI-driven robots could learn to clean up trash.
My original intention was to work on ocean stuff. I think that a database like this of stuff floating in the ocean would be awesome (and I'd love to someday collect that database), but being in Colorado makes that somewhat difficult. I think that starting with something like this on land may be a good way to understand more about what it will really take to make a dent in this area.
Comments