I assume this is picking from a predefined set of things? As in not all nouns, but a smaller list of possibilities. Otherwise it is pretty impressive that it went 6 for 6 with my terrible drawing abilities.
It has to be. I was prompted to draw "police car", and started about drawing a regular car with a siren on top. It guessed police car before I was even done drawing a generic car body - so I assume it wasn't trained on just a regular "car".
Edit: I was half wrong. It is selecting from a limited list, but both police car _and_ a regular car are actually in that list. https://quickdraw.withgoogle.com/data
I’m not sure I understand it. It prompts you to draw a certain thing, then it “guesses” what you drew, but of course it already knows because it just told you exactly what to draw. Maybe I’m missing something but the game makes no sense. If I told you what to draw then tried to tell you what you were drawing I’d get 100%.
It would be better if it just said “draw something.”
> of course it already knows because it just told you exactly what to draw
It's a computer. It doesn't know that unless it's explicitly told to remember it.
You can very easily test this by just drawing whatever you want as if it did say "draw something". It's very accurate at guessing any of the objects in the dataset, even if it's not the one you're meant to be drawing.
It'd be cool if the learning set was built on humans coming up with challenges and then strangers drawing them, and then them being guess by a third person, and then validated by the challenger as true/false.
As a few other other have pointed out, the data from this game is open-source and accessible in number of ways [1].
For an introductory exercise to deep learning for image classification, it's a great alternative (or follow-up) to the classic MNIST dataset [2], which serves as a common "Hello, world" for image-based ML.
We trained and deployed such a model for a demo of our mobile ML tool [3][4]. Feel free to ping me if you're interested, would love to chat.
I toyed with it some time ago and was impressed to see that -- in the raw data set -- each drawing is a time-series of strokes i.e. you get a person's drawing (of a penguin, the Mona Lisa, etc.) after 0.1 second of drawing, after 0.2 seconds, etc.
I made and (shameless plug!) sell giant algorithmically generated posters which illustrate and play with this idea: each group of 8 rows shows doodles after a decreasing number of seconds. It's fascinating to see how someone's drawing of a "koala" is pretty much set for "success" after a few strokes, while other drawings don't ever really improve with time... or how universal and "well-defined" are drawings of matchsticks, unlike those of kangaroos.
I too was inspired to use this dataset in some way, and made this toy program where you point your camera at your face, and it renders your eyes and nose as doodles (and barfs out more if you open your mouth): https://github.com/goberoi/sketch_face
I love using this as an English Foreign Language teacher. I just wish I could choose a subset of the vocab because some of the words are a bit obscure like "megaphone"
I drew a beautiful sea turtle; I was sure it was going to guess it! But no, it couldn't. Then I checked the training set, and it was full of tortoises...
What is amazing takes the fun away at the same time: The detection is way too fast/early. For example you are tasked to draw an octagon, the detection will already trigger once you drew three lines that are roughly aligned at 135° to each other. Draw an elephant? I can see there is something that remotely looks like a trunk and there is an attempt to draw an ear that's merely a stroke at this point. It's an elephant alright!
Things like that made me play the game in a different way: Draw it so the AI won't detect it but a human would most likely instantly recognize it. Example: Microphone. Just draw a microphone hanging from a two-segment microphone arm. Start with the arm and add the microphone last and draw a circle around it or point at it. Clearly someone would recognize this as a microphone, while the AI struggles and thinks it's something completely unrelated.
That's how I play this too! I've found it's very very bad at dealing with rotations. Carrot right side up? No problem, carrot at a 230° angle? It'll never figure it out.
I made a camera version of this a few years ago (https://www.youtube.com/watch?v=GEUrMbx_uac) using QuickDraw to train the model. I generated thousands of paper-like backgrounds (grayish noise) and pasted the raw drawings on top of it to generate JPEGs.
46 comments
[ 2.7 ms ] story [ 118 ms ] threadgoogle.com
Past discussion: https://news.ycombinator.com/item?id=12965311
Edit: I was half wrong. It is selecting from a limited list, but both police car _and_ a regular car are actually in that list. https://quickdraw.withgoogle.com/data
It would be better if it just said “draw something.”
It's a computer. It doesn't know that unless it's explicitly told to remember it.
You can very easily test this by just drawing whatever you want as if it did say "draw something". It's very accurate at guessing any of the objects in the dataset, even if it's not the one you're meant to be drawing.
https://m.imgur.com/a/y73R5Ir
I remember it worked better in 2016...
The title needs rethinking.
And, of course, I wasted half of it trying to remember what a latch circuit looked like. Before remembering they probably meant beach shoes.
For an introductory exercise to deep learning for image classification, it's a great alternative (or follow-up) to the classic MNIST dataset [2], which serves as a common "Hello, world" for image-based ML.
We trained and deployed such a model for a demo of our mobile ML tool [3][4]. Feel free to ping me if you're interested, would love to chat.
[1] https://github.com/googlecreativelab/quickdraw-dataset#get-t...
[2] https://codelabs.developers.google.com/codelabs/cloud-tensor...
[3] https://imgur.com/UagduZX
[4] https://palletml.com
I toyed with it some time ago and was impressed to see that -- in the raw data set -- each drawing is a time-series of strokes i.e. you get a person's drawing (of a penguin, the Mona Lisa, etc.) after 0.1 second of drawing, after 0.2 seconds, etc.
I made and (shameless plug!) sell giant algorithmically generated posters which illustrate and play with this idea: each group of 8 rows shows doodles after a decreasing number of seconds. It's fascinating to see how someone's drawing of a "koala" is pretty much set for "success" after a few strokes, while other drawings don't ever really improve with time... or how universal and "well-defined" are drawings of matchsticks, unlike those of kangaroos.
https://gumroad.com/l/quickdrawposter if anyone's ever so inclined!
Cheers
I too was inspired to use this dataset in some way, and made this toy program where you point your camera at your face, and it renders your eyes and nose as doodles (and barfs out more if you open your mouth): https://github.com/goberoi/sketch_face
Let me know if you'd like to do a quick collaboration!
I guess they should check their training set :P
The punctuation is important. Like "Works on Contingency No Money Down".
Things like that made me play the game in a different way: Draw it so the AI won't detect it but a human would most likely instantly recognize it. Example: Microphone. Just draw a microphone hanging from a two-segment microphone arm. Start with the arm and add the microphone last and draw a circle around it or point at it. Clearly someone would recognize this as a microphone, while the AI struggles and thinks it's something completely unrelated.
Who do you think I am?