See a Fish? Ring the Bell | Hacker News Reader

See a Fish? Ring the Bell

visdeurbel.nl

262 pointscyanbane79 comments

This excerpt from the FAQ might be interesting:

> We also want to show Utrecht’s residents and visitors how much life there is underwater in the canals. The doorbell also provides information on the species and numbers of fish travelling through Utrecht’s waterways. We can use that information to improve the quality of underwater life in Utrecht.

Of course AI could do this without human intervention. They want the public to take part in this project.

In fact, I think that's even the only true goal. Operating the gate is a manual process that takes some effort and time, so it's not like they run out to operate it every time someone pushes the button. In practice, I think they still just open it every once in a while.

As I understand it they do some scrying based on how many people press the doorbell and how frequently.

https://en.wikipedia.org/wiki/Fishfinder

  “Initially, our fish finder was sold for 600,000 yen per unit, which was as expensive as a house in those days. Despite its high price, the company’s technology attracted fishermen from all over the country. In fact, I heard that a number of customers in the fishing industry visited FURUNO (which was formerly based in Nagasaki City) with backpacks full of cash to buy our fish finder,”

  The Furuno Fish Finder is said to be the world's first practical fishfinder; it was introduced by the Furuno brothers for use in commercial fishing vessels in *1948* in Japan.

Worked for 75+ year without AI so far..

I'm not sure sonar would work all that well in a canal, given everything else in it.

It would still be good to have AI as a fallback to help the fishies out when no humans are online though...

This has been shared too many times on reddit to ever go empty I bet.

That's what they said about thiswebsitewillselfdestruct.com. https://boingboing.net/2023/12/12/this-site-will-self-destru...

Considering how still and opaque the picture is (I have just spent 5 minutes staring at the stream), you probably don't even need fancy AI. Some sort of basic movement/change detector will probably do the trick just fine. But as you say the point is to involved people.

Just as I was reading your comment, a boat went by or something and there was a sudden rush of sand and debris kicked up from the bottom of the feed, and for the past 5 minutes or so, there's been a constant stream of debris flying by and swirling around. Looks more like TV static than "still and opaque". Sometimes larger pieces of debris fly by.

But that said, nothing here that proper machine learning couldn't handle.

You could have the system flag unusual footage and ask humans (website visitors?) to review it. I am looking forward to fish captchas!

My friend, it makes up a fine entropy source, I can tell you. :)

One can always find justification for more complex solutions ;)

I think it goes both ways though. One can justify more complex solutions, only to not need the added complexity. One can always ignore edge cases to justify a more simple solution. only to find out that the cut functionality was not quite as "edge" as expected.

My security cameras love to declare that they detected motion because a car drove by and swept its headlights across my porch, and my cameras use a simple change detector.

What are the consequences of false positives? For me, I get a notification, check it, and roll my eyes. For a lock, it opens too often? I have no idea what the consequences of that are. Maybe that's perfectly acceptable and a simple change detector is enough. Or maybe it changes water balance somehow and causes ecological problems, or impacts boat navigation (I mean, the lock is there for a reason in the first place right?)

Sometimes complex solutions are justified. A camera is more complex than just looking into the water with your naked eye. But that necessitates having someone physically present, and able to see through the water. Everyone agrees that the complexity is worth it.

I've worked with too many PMs who want the fastest easy solution without regard for the actual use case and then get perplexed by the side effects (that they were previously told about but refused to grasp). And generally I wind up being the person forced to be physically present watching for fish because the PM thought adding a camera was too much complexity.

Yes, you've happened upon a common trend with new tech. Everyone wants to use their shiny new hammer, some, only know the shiny new hammer, and the wise among them know to just stick to the nails.

It happened with blockchain, it's happening with AI, it'll happen with the next big thing too, just as it has with many big things in the past.

Oh, cool idea, this should be on the (b)lockchain!

An AI could after being trained, with a higher error level than 100s of humans (or 1000s) watching. Of course, what about a trinity of AIs with quorum?

But anyhow, this sort of project.. even if announced today, likely started a decade ago in someone's mind. According to archive.org, it's been around at least 3+ years. AI wasn't realistic a decade ago, and even 5 years ago? Not really viable reliability wise.

And of course, AI isn't free... they have to be trained, security updates and software updates and hardware updates over time. I mention all of this, because everyone keeps saying "AI could do this", but could it? Because it's not just "can it do it", but "can it do it as reliably and cheaply"?

People are free. And as you say? People taking part is fun.

I'm pretty sure good enough computer vision has been around for decades. The newer stuff is object classification and facial recognition/emotional state type stuff. I think we've had good enough tech to find moving fish for quite a while.

Indeed. We have been classifying and sorting at least blueberries using neutral networks since at least the 80s or 90s.

> Of course, what about a trinity of AIs with quorum?

Caspar, Melchior and Balthazar?

I suppose we could, but I think we already know what lies at the end of that road.

> But anyhow, this sort of project.. even if announced today

It has indeed been on and off since March 2021. This project could've been set up in no time, put a camera underwater, start the live stream, and set up the basic website. Building an AI will take way longer, with much more technical skills required, and much more money.

>"AI could do this", but could it? Because it's not just "can it do it", but "can it do it as reliably and cheaply"?

Yes and also yes, ResNet was released in 2015, VGG in 2014, since then we have developed powerful models that can process a frame in a few ms even on a Raspberry Pi.

Never expected this from my hometown to hit hackernews.

For some context, here's the lock in streetview: https://www.google.com/maps/@52.0974077,5.1152216,3a,75y,277...

Not even the first time :) https://news.ycombinator.com/item?id=26645019

Utrecht really is one of the nicest places on earth

ultrech is really nice. we (my family and i) are thinking about moving to either there or hague in a few years (we are currently in ams, but we are not staying since property values are way too high).

My family tried to move to Utrecht, but ended up in Hilversum. And Hilversum is the nicest place I've ever lived for day to day life! But every time I visit Utrecht I think "if only....".

We also looked at Houten because letting my kids bike safely is a huge deal (and one Hilversum is only mediocre for) but Houten does seem quite boring...

People keep talking about training an AI to ring the doorbell but how about training the fish to do it themselves??

Why bother? Just wait a few million years and fish will develop the capacity to ring doorbells on their own.

The Dolphin Ascension will soon be upon us, and woe be upon us when Dolphinkind rise from the seas to reclaim the Earth that is rightfully theirs.

Why not train the fish to train an AI to do it?

No need, we've already trained the fish to train the humans to ring the bell.

The fish wait patiently to see if the humans will develop the technology to train an AI to ring the bell.

And the next test will be even harder.

Because water and electricity are not a good combination.

Train the fish to train electric eels to train the AI (adding indirection is both the cause of and the solution to life's problems)

This might be the first actual livestream.

It looks like this was given a second chance by mods, and moved to the front page, as it was originally posted 2 days ago, not 3 hours (at the time of writing) [0].

What's interesting is that it's the one with the editorialized title that has been rescued, despite this being against the HN guidelines.

[0] https://news.ycombinator.com/from?site=visdeurbel.nl [1] https://news.ycombinator.com/newsguidelines.html

This is probably the fifth time I've seen this posted this month.

I dread to think what the HN deathcuddle is doing to their poor doorbell.

I used to work professionally in the US doing this sort of task with computer vision. The challenging part isn’t so much the labeling/classification of a fish within an image; instead, it’s a connection to cloud environments to do processing. Most of the projects I worked on were at hydroelectric dams, where-as an ironic punchline goes—it’s punishingly hard to get access to reliable power or water.

If anyone is interested/curious happy to answer any questions on here or via DM.

Have seen this firsthand in (literal) field demos. Radios don't work. Field equipment is noisy, sound and electromagnetically. Just things. The "cloud" isn't effectively down the street. The nearest fiber is over a mile away. And so on.

The current brick wall to bang heads against is putting all of that under "connectivity" and working on that, and trying to sell solutions which are not as reliable as they should be; instead of doing compute more locally, or applying traditional system analysis to figure out what really needs to be realtime, and what "realtime" really means.

Would love to get a job in the field working on the compute / local telemetry solution somewhere in the Pacific Northwest USA. Definitionally I suppose that means somewhere you have to drive because there are no commercial flights; there's a quite lot of space like that, surprising / not surprising depends on your bias I suppose.

Well, why is it so hard to get reliable access to power or water? I can think up reasons, but you've got the background to tell me why.

Are there innovations in material science and technology that make things possible today that weren't 20 years ago? I have a shower attachment to tell me the temperature of the water, and it's powered by water through the device so there are no batteries to swap. I imagine the industrial versions must be so much more advanced than a gadget I got off Amazon.

Great question! One would intuit that the procurement function of a large organization (like one which operates a hydroelectric dam) would be able to see the value of some kind of industrial sensor that minimizes integration costs and provides them with better understanding of operations data, but practically it’s hard enough to coordinate folks to install conduit to route cables! When it requires political capital to get something as simple as cat-5 installed, clear efficiency gains easily fall off the table. It’s sort of like the infrastructure is “set in stone” and built to service the operations of the dam first, and everything else is secondary, as other commenters have suggested.

It's probably a much simpler answer. Why set up a bunch of plumbing and networking inside a dam if you don't need it? Many (most?) were built before the internet, as well. And I don't think they generally are staffed around the clock. Any more complexity than you need is a waste of money.

I bet it's exactly this. You've a massive concrete structure where adding water lines and power is going to require quite a bit of paperwork before you can even begin to consider how you run it. And they can be quite long, too, and your sensors likely need to be somewhere that power (and clean water) currently isn't.

And then you will discover just how well wifi travels through concrete whose thickness is measured in meters.

> where-as an ironic punchline goes

I knew someone who grew up a reasonable walking distance from a major operational dam, but had no grid access. They did have a generator I think, but would have had to pay the full cost of running lines to their land, and it was too expensive.

Have you tried Starlink? I work in motorsports which has a similar problem, tracks being remote and internet access being minimal, but Starlink all but solved that for us (except unfortunately in countries without service).

Late reply, but yes we had a pilot project with Starlink for a couple of locations which showed great results. Unfortunately we couldn’t use it in all cases because of stringent information security requirements; otherwise, nothing but good things to say.

Finally, an application for AI that makes sense.

My thoughts exactly. This would be a great application for it.

That said, they could also implement a bot for YouTube that allows a !ring command for when they're out of stream seats.

More like an application of Canny edge detectors

"AI", what we used to call "ML" should be able to solve this, at least 10 years ago.

ML has always been considered a subfield of AI: https://en.wikipedia.org/wiki/Machine_learning#Artificial_in...

AI as a field started in 1956, while the term “machine learning” was coined in 1959, as a sub-discipline of it.

That might be a fact in theory, but practically most companies literally did "replace all" of "ML" with "AI" in all their marketing materials and ML-engineers got rebranded as AI-engineers. ML-models are now called AI-models. Because ML is old and busted and AI is the new hotness.

Before it was called ML, it was machine vision, used in applications like quality control since at least the 80's: https://en.wikipedia.org/wiki/Machine_vision

> Every spring, fish swim right through Utrecht, looking for a place to spawn and reproduce. Some swim all the way to Germany. There is a problem, however: they often have to wait a long time at the Weerdsluis lock on the west side of the inner city, as the lock rarely opens in spring. We have come up with a solution: the fish doorbell! An underwater camera has been set up at the lock, and the live feed is streamed to the homepage. If you see a fish, press the digital fish doorbell. The lock operator is sent a signal and can open the lock if there are enough fish. Now you can help fish make it through the canals of Utrecht.

So unless there are people watching and alerting them (free labor) the fish don't get to procreate? That's a lot of responsibility.

It's so nice that they found a way to offload the externalities of running a lock without having to spend any money themselves, let the public do it. A playful implementation of "privatize profits, socialize losses". Why pay someone to actually watch the feed and react when appropriate if they can gameify and let the people do it.

> So unless there are people watching and alerting them (free labor) the fish don't get to procreate?

The lock opens roughly once per day - more often if it's busy. Once the boating season starts, fish will piggyback off of that.

It's not this hard to help fish pass a man-made obstacle. They chose to do it this way for public awareness. Leave off the public awareness goal and suddenly there are way easier, less involved methods.

This is way more for raising awareness about the impact of locks and dams on fish migration than for the actual effect of this system on fish migration.

For that goal, it works well.

> Why pay someone to actually watch the feed and react when appropriate.

There is a an operator who gets final say. This way that person doesn’t have to be driven insane with the mind numbing excruciatingly boring job of watching empty water all day.

Plus, it gets the public involved in something positive that will both entertain them and get them to think more about our impact as humans.

Sounds to me like a good way of building a high quality CV model with limited resources.

Its not apparent as an end user if the signal to the lock operator is another human or a well trained model. And at the end of the day, if there is a human checking on it, does it really matter?

I'm guessing this is a situation where the people with the power to fix this problem don't care and the people that do care don't have that power.

Would you rather they didn't do this at all?

Of course I'd rather they do it this way than not at all. It's just so sad how unsustainable it is. It's great that there is so much interest, at least right now, I could only get a read-only spot in the feed.

It's been around for years, how is it not sustainable!

I get it may not be sustainable, but is computer hardware and electricity sustainable? What about the maintenance and cost of those? This is a great solution. Beautiful. I feel like hugging this solution, as if I should bed it, and produce efficient, beautiful, socially wondrous offspring!

This reads like a very cynical take on the matter. It kind of suggests that general public should never be asked to do anything. Always hire someone for that particular thing. (help keeping cleanliness by picking random trash, report a crime or a suspicious object, paint some public walls). Some things can done better if crowd sourced and sometimes and it can be just for fun.

It is necessary to indicate that "they" are scientists, and surely service for people, not for any company.

Awesome, this is 50m from my office. Born and raised in Utrecht but had no idea.

Same, I cycle past this spot almost every day but had no idea :)

Could this not be done with ai image recognition? Also it would be great to have an interactive timeline (like a ring or nest doorbell camera) that shows previous fish and gate openings and understand busy times of day and fish species.

Edit: Someone said "why not train the fish to ring the doorbell" and that got me thinking - what if there were two chambers inter-connected with a gate - with relatively large openings on both ends like a pipe that a stream would flow through under a road... fish would swim into one side and a Sonar, IR or Conductivity sensor would know when there is something other than water in the chamber and instruct the sluice gate to open. That way the fish are opening it just by swimming in either side. It would close when nothing is detected in either chamber for a period of time. It would be the fish equivalent of walking up to an automatic sliding door at a shop.

You can still record the fish traversing and display it interactively so that people are engaged, but it doesn't become reliant on it.

Why would you automate the fun?

I see fast moving water and tons of bits of blurry things moving all about. are those fish? or debris? they could all be blurry black and white goldfish for all I know. What does a positive picture of "a fish" look like here?

It's worth pointing out one of the best youtube channels around, Monterey Bay Aquarium Research Institute

https://www.youtube.com/@MBARIvideo

Norway has done work in this area that uses computers. Their system recognizes fish and parasites. As the fish pass through the systdm a laser zaps the parasites and antibiotics are not needed.

Did anyone see a fish? I've been watching the livestream did not see any. Luckely they have a gallery with the best pictures.

I wonder how deep the camera is. There is even a picture of a bird under the water

According to this news article from the guardian about the fish doorbell the water is around 2.1m/7ft deep. Apparently the camera is at the bottom of the lock so around that depth I guess

https://www.theguardian.com/tv-and-radio/2023/jun/29/fish-do...

Also a short Youtube video from the city of Utrecht shows the whole operation https://www.youtube.com/watch?v=0MgeR85IMOM

It could be an interesting research dataset. I wonder what’s the best method today. Perhaps a fine tuned YOLO model, or a CLIP searching for fish, or some simple OpenCV?

It seems like they're saying you ring the bell when you see a fish, and that notifies the lock operator, who probably checks the same camera to verify it, then opens the locks if certain conditions are met. So, the advantage is that the operator can do other things besides watching the camera all the time, knowing that helpful people on the internet will alert him when needed.

Does that work in practice, especially after the site goes viral? I'd assume there would be a ton of false positives, i.e. people ringing the bell for the lulz.

If it's just a fun community thing, and the alert actually goes to the equivalent of /dev/null, that that's fine. Maybe a better metaphor would be those buttons on cross-walks that aren't attached to anything, but make you feel like you're influencing the light sequence when you press them. Anyway, I just don't get how this would work well in practice.

The feed is rate limited to 999 viewers. How do I know? Because I had to refresh a few times to get the doorbell UI... Imagine 900+ people ringing a "doorbell" whenever they see a fish tonight. I hope they can disable the ringing.. :D

This is amazing. They should make a little ML-operated door in the lock for the fish so you don't have to open the whole lock manually

How are the fish supposed to find it?

Wouldnt it be handier if the door bell was placed under water so the fish can reach it?

Reply on news.ycombinator.com