Multiplayer pictionary game where AI guesses your drawings. Used a subset of Google's Quickdraw dataset to trian a classification model (>75% accuracy) that is used in the AI to guess players' drawings.
Some important changes:
drawing js now only sends focusing the area drawn. maintains a square dimension. since 28x28 is also a square. trying think of a way to keep it in same aspect ratio well
using a fixed stroke weight, because variation with stroke messes up resize algorithm quite a lot
it already messes quality quite a bit when we resize some huge image into a 28x28
with my last changes, it seemed to act best
also made the backgorund pure black and stroke pure whight
white*
this is mainly to actually help the resize algorithm used
because with ur colors given when resize it blurs part and makes it into slight variaiton
so if we do things with intensity base to invert the colors, the style ends up diffuring quite a bit reducing model accuraciy significantly
but making it easy identification for bg and front, allows better resize + more suitable for the model
removd the mouse relase and send image part. because this made testing very annoying . i set it back to press enter then send image
we can switch back later. but while debug lets keep this
for now i did a few sketches. the prediction seemed reasonable. the quality still drops. but compared to the image in dataset its not that significantly worse any more
but there are still some differences that could potentially affect accuracy
noticable in my eyes were offset and the weight of strokes. like when we resize it the stroke weight gets very high with our generated images. whereas the dataset ones are not that high. but this doesnt happen always.
so update ur code and test out as much as u can
right now i have a bunch of place where I am showing the image with python
and also printing in "thoughts ..." that says what the model is predicting of the image
so use those to test things out
Some important changes: drawing js now only sends focusing the area drawn. maintains a square dimension. since 28x28 is also a square. trying think of a way to keep it in same aspect ratio well using a fixed stroke weight, because variation with stroke messes up resize algorithm quite a lot it already messes quality quite a bit when we resize some huge image into a 28x28 with my last changes, it seemed to act best also made the backgorund pure black and stroke pure whight white* this is mainly to actually help the resize algorithm used because with ur colors given when resize it blurs part and makes it into slight variaiton so if we do things with intensity base to invert the colors, the style ends up diffuring quite a bit reducing model accuraciy significantly but making it easy identification for bg and front, allows better resize + more suitable for the model removd the mouse relase and send image part. because this made testing very annoying . i set it back to press enter then send image we can switch back later. but while debug lets keep this for now i did a few sketches. the prediction seemed reasonable. the quality still drops. but compared to the image in dataset its not that significantly worse any more but there are still some differences that could potentially affect accuracy noticable in my eyes were offset and the weight of strokes. like when we resize it the stroke weight gets very high with our generated images. whereas the dataset ones are not that high. but this doesnt happen always. so update ur code and test out as much as u can right now i have a bunch of place where I am showing the image with python and also printing in "thoughts ..." that says what the model is predicting of the image so use those to test things out