yeah i agree. it is not the video how you made it but the video how you play with it. and the moment where the author says “it’s a little complicated, all you need to know is it’s small and powerful” - bro. people who watch videos with “i made an AI” in the title are usually interested in computer science and maths.
he got inspiration from ai that already can tell where you are from a single pic its like 80% accurate and its been out for like 2 years now so he didnt invent it if he did hed be able to sell it and claim intellectual rights
I agree with people above. I would add that I would at least like to see a link to GitHub page or Kaggle where the model could be taken from. It definitely looks interesting
When the square guesser gives weights on each square it believes it could be, instead of picking the center of the most probable on, what about picking the gravity center of the weights.. wonder if that would increase performance a little bit, especially when squares are clustered 🤔 Great video!
The thing with this AI is the fact that it doesn't look at a map. It looks at tiles. If somehow this AI could lock the region and then search for road layouts and types and even topographic information… The government might be at your door
This is such a great video. As someone who's working with automation and continuous improvements, I'd love to see a followup video with improvements on this. It would also be super fun to see two people try to make the best AI and compete vs each other.
This is super impressive! As someone with a programming and (some) AI background I’d be super interested to see what the process was of building the dataset, I’ve scraped small google maps/streetview datasets but it got.. pretty large, would love to hear more about your process of building it!
Out of curiousity, did you try to train a regressor on the x/y coordinate instead of a classifier? I think that it could help the system, because I suppose that it currently can be as much penalized for guessing a neighbor region as for guessing smg very far currently. Using regression would enable you to define your loss in such a way that it will penalize big errors an therefor the model will focus on avoiding this. Btw if you have a public github with your code, I would be happy to try it and make a merge request for you if it works.
Great project! I wish so much that this project becomes popular and receives incentives for improved new versions, and maybe becomes accessible to the public to play, setup, and contribute.
Yooo the video is sick. I don't know why you dont have millions of views already lol The only thing that in my mind would've improved the video is if you described the technical part in a bit more detailed way, like where you got the dataset, how did you train the AI, etc. But that's just my opinion. Anyway, you got a new subscriber ❤
It'd be great if you visualized the attention maps of your ViT network as well, I would show what the network actually looks at. I would expect it to recognize some small-scale patterns like types of plants, letters, skyline. Plus memorization of the specific camera features (like color balance, crispness etc.) used in the different parts of the world. And did you try to run it manually on the images outsize its training distribution (like for example Yandex street view panoramas or just random photos or panoramas from the internet) to check if those 96%/50% hold? That is, how much it memorized the non-generalizable features (like camera, actual season in GSV) vs the generalizable features.
I have a feeling the AI could become nearly perfect if it could also interact with the Map and this way get more context of the location it is guessing and thus perform better. It may take a lot longer to train and make predictions, but I think the boost in accuracy would be worth it.
What's more insane about this is that this AI can't even read and you spend like what, a few days training it? Just imagine a full-blown huge AI trained by like OpenAI or something across months of non-stop training. It would probably be able to get the exact coordinates within like a few meters.
I’d like to see some kind of graphic made from this AI that shows some kind of similarity between different blocks or countries. I wonder how much it picks up on camera quality.
A bit disappointed about the title and what was actually in the video. "How I made the best Geoguesser AI in the world" -> "It's complicated, I just put 600 000 images to the model and trained it 18 hours." .. Yeah, ofc you put a shitload of images and their locations, but that's the part that is interesting, but 3/4 of the video was about playing against it. I really hope you're going to release a more detailed video. What hardware, libraries,, what was working during your test, what wasn't, data normalization, finetuning on geoguesser meta ( poles, roads, cars etc ), how you created the "region grid"All the interesting stuff is missing ! 😁
Even if you don’t make a video on it, would really appreciate some sort of public source code on this. Myself and others have spent a lot of time on Geoguessr AI and this could be interesting to learn from.
Wow, this is really impressive! What language did you use to send the coordinates to Geoguessr automatically? It could be improved by moving into a different country when it selects a country without coverage or a road with coverage. However, if it's between 2 roads with coverage, I think it could just hedge. Knowing the road angle could also help, I think it wasn't scanned in the input. This is a really good AI, I don't think I can make something remotely impressive as this though, and this is your first video!
What dataset are you using? Did you use the maps API or is there one available? I would be very interested in the code and further details. Btw great video
what about a more algorithmic predetermined approach for a bot to play it? like recognizing tree species or flowers and cutting the area down to a smaller radius?
Wow great video and great project. I am just wondering how did you collect data for training model and how did it take so little time to train it. And do you have some tips for someone starting with ai?
albo to przypadek ale grałem z kimś z ameryki z podobnym skinem nie pamiętam najpierw źle trafił ale resztę gry dominowal klikajac prawie idealnie wszędzie
nice would be to change a bit the AI brains to be able to get out useful insights in an automated way of what to look, and what characteristics the AI is using to guess the region. I bet there is dumb stuff that we overlooked that help to guess right. I would love to learn this stuff.
I didn't hear you say anything about separating you training and testing sets. Without doing that, the AI may just have "memorized" the images it was given during training, and it may not be able to generalize at all. I'd definitely want to hear more about the design of the AI.
Hi, I'm currently doing a research project at University and this video is very fascinating. I would like to know how you get the training dataset and set up the bot to interact with the game. Also if you can link me some of the research paper, that would be great
Do you know to what extent minor image imperfections like humanly imperceptible smudges on the lens are used by the AI? Do you have a way of illustrating which part of a picture provide the most information?
![Felix "Unfair" | [Stray Kids : SKZ-PLAYER]](http://i.ytimg.com/vi/Oswujxm2Ag0/mqdefault.jpg)
Would love to see a more detailed video on how you actually created it/how it works in the background
yeah i agree. it is not the video how you made it but the video how you play with it. and the moment where the author says “it’s a little complicated, all you need to know is it’s small and powerful” - bro. people who watch videos with “i made an AI” in the title are usually interested in computer science and maths.
Yes, please
he got inspiration from ai that already can tell where you are from a single pic its like 80% accurate and its been out for like 2 years now so he didnt invent it if he did hed be able to sell it and claim intellectual rights
I agree with people above. I would add that I would at least like to see a link to GitHub page or Kaggle where the model could be taken from. It definitely looks interesting
Damn really impressive
THIS IS YOUR FIRST VIDEO?
Thanks GabrielGeoo for taking me here
Same
When the square guesser gives weights on each square it believes it could be, instead of picking the center of the most probable on, what about picking the gravity center of the weights.. wonder if that would increase performance a little bit, especially when squares are clustered 🤔
Great video!
this is amazing stuff man, great video!
This is a banger video and deserves more recognition
I am not intrested in AI or Geoguessing but the video was great and I am suprised it hasn't millions of views.
me to
21k views for 600 subs isn't bad
The thing with this AI is the fact that it doesn't look at a map. It looks at tiles. If somehow this AI could lock the region and then search for road layouts and types and even topographic information… The government might be at your door
They already do that, it’s called your address and it locates your exact region and road layout
What impressed me the most about this video was your ability to quickly recognize the Korean power pole at 5:42
This is such a great video. As someone who's working with automation and continuous improvements, I'd love to see a followup video with improvements on this. It would also be super fun to see two people try to make the best AI and compete vs each other.
Now this is the content i want to watch, high quality and educational on how to solve a problem
Great video, can't believe it hasn't blown out yet
Hi !
Where did you get the training data ?
Are the images used for the training the same frames as the ones that are used on the geogussr maps ?
I would guess he performed a massive scraping of google maps locations
@@hadriencrassous2162 I think geogueser use the same google maps images data. Maybe thats why the model works so well.
@@ariefwt2220 probably able to basically get 100% with a big enough model that simply learns all the pictures
Truly amazing content. Keep it up!
I was quite literally going to look into doing this myself and you beat me to it. Props to you!
Great video! Always excited about new AI content creators. Hoping to see some more exciting stuff soon :)
Impressive work, keep going!
amazing video!
This is super impressive! As someone with a programming and (some) AI background I’d be super interested to see what the process was of building the dataset, I’ve scraped small google maps/streetview datasets but it got.. pretty large, would love to hear more about your process of building it!
How is this your first video! The quality is outstanding!
Awesome vid bro 👊🏻
Out of curiousity, did you try to train a regressor on the x/y coordinate instead of a classifier?
I think that it could help the system, because I suppose that it currently can be as much penalized for guessing a neighbor region as for guessing smg very far currently.
Using regression would enable you to define your loss in such a way that it will penalize big errors an therefor the model will focus on avoiding this.
Btw if you have a public github with your code, I would be happy to try it and make a merge request for you if it works.
Yeah, I was thinking the same thing!
You can do distance-based label smoothing to not penalize the model too much if it's close to the real position.
@@gaggix7095 lol once again exactly what I was thinking as well
Regressions were found to have subpar performance years ago in papers, which is why they moved to classification/retrieval instead
@@larrygan9839 L1/L2 loss is perfectly fine when used on the correct task
this is amazing! please go more in depth how this ai operates :)
Great project! I wish so much that this project becomes popular and receives incentives for improved new versions, and maybe becomes accessible to the public to play, setup, and contribute.
great content
Really neat to see the transformer pick up on geo-specific hierarchical features so well
I thought I was watching some big youtuber who has a name in this field, surprised to see you are not, and wish you all luck and success!
This is an amazing video. I don’t understand how you have so few views.
Yooo the video is sick. I don't know why you dont have millions of views already lol
The only thing that in my mind would've improved the video is if you described the technical part in a bit more detailed way, like where you got the dataset, how did you train the AI, etc.
But that's just my opinion.
Anyway, you got a new subscriber ❤
release it open source and spark a community I really wanna see how far this gonna go ♥
y geoguesser se va a llenar de hackers haciendo puntuaciones perfectas 😅
That's a horrible idea, GeoGuessr will be ruined.
Nice vid! Looking forward to the content :3
It'd be great if you visualized the attention maps of your ViT network as well, I would show what the network actually looks at. I would expect it to recognize some small-scale patterns like types of plants, letters, skyline. Plus memorization of the specific camera features (like color balance, crispness etc.) used in the different parts of the world.
And did you try to run it manually on the images outsize its training distribution (like for example Yandex street view panoramas or just random photos or panoramas from the internet) to check if those 96%/50% hold? That is, how much it memorized the non-generalizable features (like camera, actual season in GSV) vs the generalizable features.
Really good video man, keep it up 👍
That's amazing work, keep up
Honestly. Compared to the amazing video quality. I am surprised this has not gone viral. anyway. keep it up!
its very imprrssive that something this simple can perform this good
I have a feeling the AI could become nearly perfect if it could also interact with the Map and this way get more context of the location it is guessing and thus perform better. It may take a lot longer to train and make predictions, but I think the boost in accuracy would be worth it.
Imagine making this a tool for reverse searching. Users can upload a photo and know exactly where it was taken
uh yeah this isn't concerning in the slightest and definitely cant be used maliciously
Thankfully only this guy has access to it.
What's more insane about this is that this AI can't even read and you spend like what, a few days training it? Just imagine a full-blown huge AI trained by like OpenAI or something across months of non-stop training. It would probably be able to get the exact coordinates within like a few meters.
that is amazing also how is this your first video while also being an amazing video
how do you only have 700 followers ? Very cool video
Amazing video, would love to see a more technical video! Open source ???
This is really cool!
Only 45 subs??? First 50, remember me when you blow up!
^
Pretty sure a group of students from Stanford built a model for this that was extremely accurate.
I’d like to see some kind of graphic made from this AI that shows some kind of similarity between different blocks or countries. I wonder how much it picks up on camera quality.
It's insane!
Great stuff!
This is amazing content! How do you only have 50 subscribers??
Good luck with your youtube carrier, it's peak
A bit disappointed about the title and what was actually in the video. "How I made the best Geoguesser AI in the world" -> "It's complicated, I just put 600 000 images to the model and trained it 18 hours." .. Yeah, ofc you put a shitload of images and their locations, but that's the part that is interesting, but 3/4 of the video was about playing against it. I really hope you're going to release a more detailed video. What hardware, libraries,, what was working during your test, what wasn't, data normalization, finetuning on geoguesser meta ( poles, roads, cars etc ), how you created the "region grid"All the interesting stuff is missing ! 😁
Exactly
The Toxic Avenger soundtrack is a perfect fit for this video 🤩
and Vivaldi lol
Very impressive
Even if you don’t make a video on it, would really appreciate some sort of public source code on this. Myself and others have spent a lot of time on Geoguessr AI and this could be interesting to learn from.
We need an ai vs rain bolt
Just save the photo and location every time it gets confused then train it a bit more on just those places once you got maybe 50 stored up.
Wow, this is really impressive! What language did you use to send the coordinates to Geoguessr automatically?
It could be improved by moving into a different country when it selects a country without coverage or a road with coverage. However, if it's between 2 roads with coverage, I think it could just hedge. Knowing the road angle could also help, I think it wasn't scanned in the input.
This is a really good AI, I don't think I can make something remotely impressive as this though, and this is your first video!
I would be very intrested in a deep dive on how you built all the project 🎉
How does this compare against the standard PIGEON GeoGuessr AI? The claim of best geolocation AI is a bold one.
i want to see rainbolt vs this ai now
Awesome video! Do you have a computer science background? :)
Super cool! Maybe you could give it the ability to move around?
I would love a way more in depth machine learning vid on this plz 🙏🙏
Does it work on photos that aren't on geoguessr? I wonder if it's just ended up memorising all of the images in the game
Marking down 185 subs for now- cuz I’m sure that will go up a lot
Love it !!
Ur underrated
This one is gonna be a viral
What dataset are you using? Did you use the maps API or is there one available?
I would be very interested in the code and further details.
Btw great video
Asking me the same
He might have scraped it himself using the Maps API
If this thing could scan for road direction it could be reliably getting 5ks
Really small but really powerful. Sound familiar
That one country he did not train it for
I wonder if this can be used to develop new metas like how exactly did it region guess Peru what did it see that we might not have
Nah bro his montage is a 1m + youtuber and he have 1k only 😮
Isn't this a clear case of train data leakage during test as GeoGuesser uses Google Maps data just like your training data?
Probably yeah. Would be interesting to test out, whether a place that is not in the streetview dataset could be recognised accurately.
Guy, I need this AI, share a link to it plz
This cool do you have a blog or a github related to this it would be interesting to check out.
so awesome!
get rainbolt on it boys
I wish you'd explain the region guessing part a bit better, what is it looking for? how does it work? what specifically did you train it on?
what about a more algorithmic predetermined approach for a bot to play it? like recognizing tree species or flowers and cutting the area down to a smaller radius?
not bad!
Wow great video and great project. I am just wondering how did you collect data for training model and how did it take so little time to train it. And do you have some tips for someone starting with ai?
albo to przypadek ale grałem z kimś z ameryki z podobnym skinem nie pamiętam najpierw źle trafił ale resztę gry dominowal klikajac prawie idealnie wszędzie
nice would be to change a bit the AI brains to be able to get out useful insights in an automated way of what to look, and what characteristics the AI is using to guess the region. I bet there is dumb stuff that we overlooked that help to guess right. I would love to learn this stuff.
is the region guesser just a one-hot classification layer on top of the ViT?
Crazy
Would really love to play with this. Can you make it open source?
You should release the model weights
Bruh, the first place was literaly 30km away from me :D
Keep up the good work impressive af
Duude awsome
I didn't hear you say anything about separating you training and testing sets. Without doing that, the AI may just have "memorized" the images it was given during training, and it may not be able to generalize at all. I'd definitely want to hear more about the design of the AI.
Instead of using squares, could we use semantic geocell creation to leverage regional distinctions?
whats the accuracy difference without any training?
Have thou thought of using the YFCC100 dataset in addition to google maps?
Hi, I'm currently doing a research project at University and this video is very fascinating. I would like to know how you get the training dataset and set up the bot to interact with the game. Also if you can link me some of the research paper, that would be great
Do you know to what extent minor image imperfections like humanly imperceptible smudges on the lens are used by the AI? Do you have a way of illustrating which part of a picture provide the most information?
This is awesome. What dataset did you use? I can't find any with 600k samples online