（评论）

（评论）
(comments)

原始链接: https://news.ycombinator.com/item?id=40767459

在本文中，用户想象了一个普通的一天，尽管城市采取了安全措施，但仍有一个物体意外从上方坠落。他们表达了担忧，尤其是那些有恐惧或健康问题的人。然而，他们澄清说，上述技术不涉及瞄准或瞄准，其机器学习功能只是识别物体。该系统不适用于空中轰炸，因为所涉及的复杂性远远超出了其能力。作者批评了不必要的政府干预，并分享了一个展示不同技术成就的 YouTube 链接。他们提到了写作过程中面临的挑战，可能暗示了机器学习和机器人技术方面的专业知识。关于视频对象检测，用户想知道为什么需要 Roboflow SaaS，而不是 Frigate 或 DOODS 等本地解决方案。他们要求评论与 Roboflow 进行比较，暗示潜在的性能优势。最后，他们讨论了 Pi 的推理速度慢，并指出了诸如访问大量预训练模型、支持最新 SOTA 模型、与数据集/注释工具集成以及对下游任务的良好支持等优点。

It seems I'm in a minority thinking this is not that great... wind can blow the hat (or the thing from the generalized idea) into traffic, or onto a baby, or any other place to upset people. Also, if the recipient can't/doesn't pick the thing up, then it's littering. From the technical perspective finding heads in a video is not that impressive nowadays... So, I don't get all the excitement...

Case in point - my friends wanted to imbibe a certain white powder with alcohol and I had to let them know that it is magnitudes more toxic to take them together. Did they have a less fun time? Probably. But I won't have their premature deaths on my conscience.

So, now we're comparing a hat dropping one or two stories to overdosing on a combination of drugs and alcohol?

It'd be nice if we could all just chill the hell out and let someone's fun, stupid, kinda pointless project just be someone's fun, stupid, kinda pointless project.

Edit: I don't mind the downvotes, but do feel free to tell me if I'm off base for thinking comparing this project to overdosing is a hell of a stretch.

Imagine walking (or in this case, standing around) on a sidewalk just going about your business. Then, imagine something drops on your head, literally out of the blue. In a city littered with scaffolding designed to prevent pedestrians being injured by stuff dropping from buildings. Further, imagine you are easily scared and/or have a weak heart. So, I think it's not a huge stretch to say that, with enough unlucky coincidences, this also might kill someone.

I was responding with an anecdote to the comment that sometimes it's important to communicate concerns, even if it means being a killjoy. Didn't mean for it to sound like I was trying to equate the AI hat-dropper to potentially overdosing, just a recent occurrence that I was reminded of when I saw the parent comment.

I don't think he was making that comparison. I think this he was more referring to the mentality of "you must be fun at parties" whenever someone speaks up with some concerns about an idea.

Eh, it's cute and I seriously doubt he is using this when not at home since it is a single use device.

But also I am shocked that there is a New Yorker that would pick up a hat from the street and put it on their head. My first thought would be, "how much lice is in this thing?"

I was once on 25th street in Midtown, when I saw someone drop a tiny object with a little parachute from a window at least 8 or 9 stories up. Once it had finished slowly gliding down to the street, someone picked it up and used it to enter the building. It was the key - I guess the buzzer didn't work! It was a delightful sight.

I've seen people use an electric whisk to unreel a string to which the key was attached. Seems less risky than the parachute thing.

It must have been dramatic, but from a practical point of view wrapping it inside a few layers of paper towel (so that it doesn't kill anyone) would be faster, and easier to target.

Is it even possible for a a parachute-retarded key to directly hurt someone? I’d be more worried about it surprising someone driving a car or riding a bike and causing an accident

I think the GP meant dropping the key wrapped in paper without the parachute. That's what makes it faster and easier to target. Which also answers your question about surprising someone driving a car, since the key won't drift with the wind anymore.

I wouldn't say it's impossible, but when you're driving/biking in dense, busy cities, you encounter all sorts of unexpected dangerous fast-moving obstacles all the time. I can't imagine this would be a bigger problem than any of the other random shit flying around Manhattan at any given moment.

if i am curious about something and want to learn, i don't want to need to sift through jokes and sarcastic comments. i find joy in learning and people can still be informative and use humor.

I can’t wrap my head around how that hat drops in a straight line. Between the propeller and any wind, how is that hat not all over the place?

If you watch the video, it actually falls several sidewalk tiles away and he has to go pick it up. From the text of the blog, I had assumed he was using AI to actually land it directly on a person’s head, which would’ve been crazy impressive.

OpenCV was not the "AI" here, the "AI" was a computer vision model trained at the roboflow website that he mentioned multiple times and that he used in the line commented with "# Directly pass the frames to the Roboflow model".

I mean, the site is pretty blatant viral marketing for both his drop-shipped-hats-from-china side hustle and (I'm going to go out on a wild limb here and guess) his employer's ML-dataset-management-related startup.

I wish cool stuff like this wasn't always sullied by the slimy feeling from it only being done to draw attention to some startup sitting smack in the middle of the trendiest buzzwords of the month.

I can assure you that if you develop a system to accurately place objects (bombs, say) on top of people and post the code on the open internet for everyone to see, the government will indeed have some critical question for you.

Accurately placing heavy, aerodynamic objects onto people when you start out directly above them is not very difficult. The hard parts are either placing the object on top of the person from a few hundred or thousand miles away, or - in this case - placing an object that tends to flutter rather than follow a ballistic trajectory.

> Accurately placing heavy, aerodynamic objects onto people when you start out directly above them is not very difficult.

It's still difficult; to do that, you need to know the wind speed at every point between them and you.

Or you need to be so close that the wind speed doesn't matter, but at that point nobody's going to be impressed that you can hit them.

I can assure you that you have no idea what you're talking about, starting with the fact that you obviously didn't watch the video.

It isn't aiming anything. It isn't adjusting for anything. It's doing so from a stationary point.

The ML isn't used for anything other than a simple "is there the thing I was trained to look for within this area?" It's basically a ML version of something one could pretty easily do in OpenCV.

There's NOTHING about this useful for aerial bombing, which involves dozens of problems much harder than "this is the spot you should aim for."

There are probably dozens of smartphone apps for helping marksmen calculate adjustments that are about a hundred times more complicated, and more useful for (potentially) hurting people, than this.

And then there's this Stuff Made Here project where the guy makes a robotic bow that can track objects and hit them no matter where you're aiming: https://www.youtube.com/watch?v=1MkrNVic7pw&pp=ygUTYm93IHRoY...

I can't stand people who act like it's reasonable for the government to monitor and harass people for stuff like this. The second our government is harassing him or the SMH guy, I'm moving to Canada.

You've replied to somebody talking about "if somebody developed (something not in this blog post)" with a long angry rant as if they had imagined the blog post claimed it had developed that thing.

I was disappointed by that, too.

Now if you had terminal guidance... Put flaps on the hat, and use shape-memory alloy wire and a coin cell to actuate them. The hats follow a laser beam projected by the drop unit. Minimal electronics required in the hat. This is how some "smart bombs" work.

It looks like it has more to do with the aerodynamics of the hat than the wind. It also hits a ledge on its way down in the video.

It seems like both of these are tractable issues.

A round hat that is spun with a significant initial angular momentum would probably fair better in landing more predictably.

> Imagine using AI to drop an object and it falls perfectly where you want it.

There is a fantasy series that depicts this as a game that two young gods would play together when they were growing up. (Or rather, since one of them had vastly superior foresight to the other one, he'd bully his brother into playing with him.)

I love this kind of project.

A lot of states are working on legislation that includes requirements for watermarking AI generated content. But it seldom defines AI with any rigor, making me wonder if soon everyone will need to label everything as made with AI to be on the safe side, kinda like prop 65 warnings.

This is not quite like the "AI" that's hyped in recent years, the key component is OpenCV and it has been around for decades. Few years ago, this might have been called Machine Learning (ML) instead of Artificial Intelligence (AI).

So it doesn't actually drop hats onto heads and doesn't use what most people would consider AI... I think I could probably rig up something to gracelessly shove an item out of an open window too which is basically what we're left with. It'd take longer to create the app for booking appointments, and to set up everything for payment processing.

You have discovered a secret area of my personalized "pet peeves" level: just a few days ago I saw an article (maybe video) about how "AI" tracks you in a restaurant. Screenshot was from an OpenCV-based app with a bounding box around each person, it counted how many people are in the establishment, who is a waiter and who is a customer, and how long they have been there.

There's an old saying: "Yesterday's AI is today's algorithm". Few would consider A* search for route-planning or Alpha-Beta pruning for game playing to be "Capital A Captial I" today, but they absolutely were back at their inception. Heck, the various modern elaborations on A* are mostly still published in a journal of AI (AAAI).

This is a fair point and maybe someone more well versed can correct me but pretty much all state of the art image recognition is trained neural networks nowadays right? A* is still something a human can reasonably code, it seems to me that there is a legitimate distinction between these types of things nowadays.

Yeah sorry, rereading, that came off as way aggressive for no reason. Rereading the chain, I think I just meant that it’s an algorithm that was frequently taught in AI classes, so at least some profs think it counts, even though it was called an algorithm.

same as parent, it was taught to me in an introduction to algorithms class, and no one during my academic stay ever referred to it as an AI.

I don't disagree that it certainly meets certain AI criteria, just saying that particular phrasing (A* is AI) was never used.

Yes, no more machine code. Everything was to be written in BASIC. ...how we laughed at that outlandish idea. It was so obvious performance would be... well... what we have today pretty much.

If you spend a few hours writing a bit of code that has to run for decades, millions or billions of times per day on hundreds of thousands or millions of machines it seems quite significant to use only the instructions needed to make it work. A few hundreds of thousands extra seems a lot. One would imagine other useful things could be done with quintillions or septillions of cycles besides saving a few development hours.

Maybe it is easier to define what isn't AI? Toshiba's handwritten postal code recognizers from the 1970s? Fuzzy logic in washing machines that adjusts the pre-programmed cycle based on laundry weight and dirtyness?

Be it "AI" or not, these mostly fall under "AI" legistlation, at least in the new EU AI Act. Which is IMHO a better way to legislate than tying laws to specific algorithms d'jour.

If Big AI lobbyists get their way, this is exactly the kind of warnings we'll get.

Flood users with warnings on everything and it'll get ignored. Especially if there's no penalty for warning when there isn't a risk.

Big Tobacco must love Prop 65 warnings, because by making it look like everything causes cancer, smokers keep themselves blissfully ignorant at just how large the risk factor is for tobacco compared to most other things.

I fear you’re right — cookie banners will soon also come with endless AI disclaimers that net net desensitize the end user to any consideration as they seek to skip poorly crafted regulation and get on with their lives.

Poorly enforced regulation. Most of the cookie banners are illegal but businesses, especially large ones, have too much power to be effectively regulated.

The nags are kind of malicious semi-compliance, partly in effort to make the regulation look bad.

This comment is known to the State of California to contain text that may cause you to ignore warnings which may lead to cancer, reproductive defects, and some other shit that I can't remember because it's been almost a decade since I lived in California and weirdly I can't easily find the full text of one of these online through a quick search (emphasis: quick)

If the goal is to make a window-based store, then why do you need AI at all? Just release the hat once payment goes through.

This reminds me of thousands of blockchain projects that used the technology to flip on light switch.

I believe the whole project, and the talk of stores in particular, is humour. At least that's how I read it. I appreciate not everyone has the same sense of humour so that may have passed you by.

This concept is great, it’s also a brilliant idea for a webcam on a Bourbon St balcony in New Orleans to throw beads at parties below. I am friends with a guy who owns a multistory bar in the middle of the strip and would be open to this, so if OP or someone else is interested in developing an AI/remote control bead thrower, drop some contact info and I’ll reach out

I am seeking neighboring stores! Sometimes I crave gum on the street, Gum drop anyone?

To summarize, I used:

1. Low weight but very cool product (like Propeller Hats)

2. Raspberry Pi for controlling everything

3. Adafruit stepper motor for the dropping mechanism

4. Yarn for holding the hat

5. Roboflow for the AI

I dream of a world where I merely open my mouth and wish it and the gum just flies down into it, already unwrapped.

You’re working toward this world and I commend you.

I would hope that we have invented error-free software development by then, though. Otherwise, a small error leading to the wrong coordinates could really ruin your day (or head)... ;)

Or use lasers and tiny gum-shaped smoke bombs to sample and model the local air column currents, pre soften and flatten a portion of the gum paper-thin with some sort of wettimg/rolling assembly, stage, then let it drop and form its own miniature gum parachute or replica of one of those whirling propeller seeds that have a built-in wing to slow their fall.

What about a “we will remember it for you wholesale” version of the gum experience - you pay money and are then implanted with memories that are indistinguishable from chewing the gum. I kinda think this is the end goal for all capitalism - you pay money for nothing.

Apparently the knowledge isn't wide enough, because this is the first I'm hearing of it... Why is gum bad for you? I knew it was in a downward sales trend, but I figured that was just consumer preferences changing over time.

Why does this remind me of something out of a certain old point and click adventure game, it was one that had the verb USE apply to every type of action.

click>(GUM)

click>(SELF)

click>(USE)

"You used the GUM on yourself.

Nothing special happens.

You now have 0 GUM."

There was another game in the same genre that did the same, but with the verb OPERATE. As teenagers my friends and I used to laugh way too much at dialogue responses these games would craft, where you would get things like "OPERATE GUM on SELF"

Maybe a receiving chute? Small, portable, and a clearer indication (cannot be confused with a yawn), plus it'll open up the variety of comestibles you can purchase just s mouthful of. No more forks, no more spoons, just a little sloped thing to slow and guide

What if we had like a fridge with glass window and drinks or snacks organized in rows with identifiers for each. You could enter the identifier and make your payment to the fridge and it would drop the corresponding drink/snack to a slot on the bottom of the fridge.

i work on roboflow. seeing all the creative ways people use computer vision is motivating for us. let me know (email in bio) if there's things you'd like to be better.

Slightly unrelated: Did the building owner/landlord complain about that? Is it legal?

I know a friend of mine whom the building asked to remove a camera they had. It was a camera used only to record the hill view in front of the building, so it isn't violating any privacy, and it was attached with magnets, so no damage whatsoever.

I was also curious about this. a bunch of BASE jumping hats dropping off a building is exactly the sort of project I would momentarily think about doing and never seriously entertain due to being certain that sooner or later someone, somewhere is going to sue me for some marginally harm-like side effect.

I don't know how litigous your region is but of all the people you know who have been sued, how many of them got sued for something silly vs a more low effort scheme like the classic throw yourself onto someones car and have 'back pain'? You might be safe to do silly shit on the basis that there are easier and better targets available.

Also curious if they had any grounds for that. I was under the impression that if you have a camera within your apartment (looking through window), nobody should be able to tell you no.

Unless perhaps the camera was attached outside their window (no longer their apartment), in a way that could be deemed unsafe and fall off and hurt someone, whereupon the building owner could be held liable? In that case I would find it reasonable to tell them to remove it.

> Unless perhaps the camera was attached outside their window

I remember it was on the balcony, securely attached. The building simply cited their policy, not any laws nor safety issues.

> Sometimes I crave gum on the street

My immediate response to this was “ew, there’s already so much gum on the street”. Then I realized you meant you want to chew gum while walking down the street and I became enlightened.

What an unexpectedly cool post, I clicked the link thinking it would be "typical dumb", but it ended up being atypically dumb in the greatest way! Fascinating. The author overcame many challenges and wrote about them in a style as if he solved the hardest parts with only a little fiddling. Maybe he's already seasoned in the ML and robotics domains? So much fun to read.

Regarding the Video Object Detection:

Why does inference need to be done via Roboflow SaaS?

    ...(api_url="https://detect.roboflow.com", api_key="API_KEY")

Is it because the Pi is too underpowered to run a fully on-device solution such as Frigate [0] or DOODS [1]? And presumably a Coral TPU wasn't considered because the author mostly used stuff he happened to have laying around.

Can anyone comment contrasting experience with Roboflow? Does it perform better than Frigate and DOODS?

Asking for a friend. I totally don't have announcement speakers throughout my house that I want to say "Mom approaching the property", "Package delivered", "Dog spotted on a walk", "Dog owner spotted not picking up after their beast", and so on. That last one will be tricky to pull off. Ah well :)

[0] https://github.com/blakeblackshear/frigate/pkgs/container/fr...

[1] https://github.com/snowzach/doods2

You are hereby put on notice that the undersigned intends to and henceforth will appropriate for his own further use without attribution to you the phrase “atypically dumb in the greatest way,” and furthermore that the undersigned may modify said phrase by replacing “greatest” with “best.” Any objection by you to said appropriation and/or modification by said undersigned will be and thereby is deemed waived by you, provided you do not respond to this notice within 48 hours. Please redirect your reply, if any, to /dev/null. Thank you.

FWIW you can use roboflow models on-device as well. detect.roboflow.com is just a hosted version of our inference server (if you run the docker somewhere you can swap out that URL for localhost or wherever your self-hosted one is running). Behind the scenes it’s an http interface for our inference[1] Python package which you can run natively if your app is in Python as well.

Pi inference is pretty slow (probably ~1 fps without an accelerator). Usually folks are using CUDA acceleration with a Jetson for these types of projects if they want to run faster locally.

Some benefits are that there are over 100k pre-trained models others have already published to Roboflow Universe[2] you can start from, supports many of the latest SOTA models (with an extensive library[3] of custom training notebooks), tight integration with the dataset/annotation tools that are at the core of Roboflow for creating custom models, and good support for common downstream tasks via supervision[4].

[1] https://github.com/roboflow/inference

[2] https://universe.roboflow.com

[3] https://github.com/roboflow/notebooks

[4] https://github.com/roboflow/supervision

> ... "Dog spotted on a walk", "Dog owner spotted not picking up after their beast", and so on.

How about hanging a London Tube-style yellow dot-matrix display showing estimated times of neighbours walking past your home? Something like:

"1. Mrs Green towards Post Office 5min"

"2. Mr Smith towards Bus Stop 7min"

"3. Mr Snow towards Mrs Smith 9min"

Love this! I play recreational ice hockey in an Adult league and for the past many years I've desired to use AI/Object recognition to recognize who was out on the ice during what times during the game to attribute who impacted goals and which players were taking longer than usual shifts ( every team has those one or two players!).

This may be achievable for me with the current state of AI and GPT to help fill the gaps that my knowledge is lacking in. Thanks for showing what you made and how you did it. It's encouragement to me.

This would be interesting, feel free to email me if you get stuck. If you had a camera at eye level, you could try to train it on recognizing the player jersey numbers.

Facial recognition would be better. Don’t forget that canonically in Mighty Ducks D2 Goldberg and Russ switched jerseys so that Russ could get his infamous “Knuckle Puck” shot off undisputed because everyone thought the puck was passed to Goldberg until the mask came off. So the ML training on jerseys would have missed this critical moment and potentially assigned the score to Goldberg, when really it was Russ (wearing Goldberg’s jersey) who should have gotten the credit.

One might argue that this sort of thing rarely happens so it’s not worth doing more complex facial recognition vis a vis Jersey numbering. But I say that while it may be rare, when it does happen it’s a major event, so no complexity should be spared to ensure we capture it accurately.

I play in a rec soccer league and had a similar idea, except to also have everyone on the team wear a smartwatch that could intelligently buzz at you to sub out based on your heartrate and how long you've been in.

should give this to the coach too - Texas players get heat exhaustion

Trace and hudl use shirt number and person tracking. I bet they could add skin color and gait analysis to do this as well.

> Picture a world where you can walk around New York City and everything you need is falling out of windows onto you. At a moments notice, at the drop of a hat. That's a world I want to live in. That's why I'm teaching you how to do yourself. Remember this as the first place you heard of "Window Shopping."

I truly love the concept of pun-driven development (PDD). As a motivating economic principle, a world where every human being has the resources, time, and personal safety to dedicate absurd amounts of their time to inane levels of pun-driven development is perhaps my favorite definition of utopia.

It can't be the best. It's only one of many positive consequences. Not even a main justification, but only a point of defense for those so irrationally against the concept.

This is probably a bottom of the barrel idea if you took it in that world where everyone can experiment and execute their ideas. Like, this would probably get you put in jail in that world, it's that lame.

Sometimes I feel we live in a simulation in a real world a few levels down with universal income or something like that. They got bored so had to forget their existence by creating a simulation (or nested simulations).

It would be cool to make something similar for a pet feeder. Imagine having two cats (like we do). A skinny one and a fat one. AI would recognize them and dispense more food for the skinny one throughout the day. Hmm... :-)

I feel like such a killjoy, but the first thing I thought of is the ongoing lice “epidemic” among people with school aged children in NYC.

I have never liked it when the ACs drip on me in midtown let alone a hat dropping on my head!

Although I think the idea of nonconsensual hat drops is so fun and fantastic.

I wish I could register myself as being up for any sort of serendipity like this. While I like the idea of a hat randomly dropping onto my head, some people may not.

Most underrated comment!

Also, this would be contrary to GP's comment - it would be the right reason. Imagine if a bald person is walking by and a toupee happens to fall on their head and they can see themselves in a window reflection of a toupee shop that just so happens to be there.

Use some ML/AI to choose the right fit, style, hair color etc., the drop orientation, and angle. Throw in some ChatGTP integration to suggest using scalp glue. Combined with OP's marking skills they will be in business in no time!

This is so cool and just brings me a lot of joy :)

Also, I've been working on a project (non-commercial) that looks down on people and have found existing models don't work super well from that angle so thank you for publishing your work on Roboflow.

Fantastic, I love this kind of silly stuff. The clear next iteration is a 4-prop hat, which can be guided to the target head.

Of course, that starts to verge on what's spooky about the idea, but either way, this is really fun and cool.

This is cool. It reminded me of a dream project in my backlog: I want to build a fan that tracks my head when I workout and always blows at my face.

Do y’all think a similar stack/setup (raspberry pi and python3 and this model thing he linked to) would be a good starting point? I prefer to use a more “algorithm” solution than a full blown model (I mean cameras have had face detection since what, the early 2000s?).

Anyway, curious to hear any suggestions.

Love the creativity and humor which is often the spark for true innovation.This guy is a real life Kramer from Seinfeld. Reminds me of the episode where Kramer drops a ball of oil from his nyc apartment while testing a business idea.

> Picture a world where you can walk around New York City and everything you need is falling out of windows onto you.

A funny way of criticizing something. Great commentary.

Once superintelligence takes over all jobs, as it is claimed will happen (, and there is an AIBI : AI Basic Income), I hope we are free to do more such projects :)

> My dream is for all the city windows to be constantly dropping things on us all the time. You will need a Raspberry Pi...

A Raspberry Pi would hurt quite a bit, depending on the floor!

Can you go a bit more in depth for the part regarding training the Ai to recognize the heads? Like what software(s) did you use ecc... I'm an undergrad who's seeking to do similar computer vision internships for his thesis and I find this kinda fascinating

（评论） (comments)

（评论）
(comments)