I ran the image regression problem where discussed in lesson 6. Then I tried to predict using some random images from the internet.
Then I got this:
I think our model is optimized for that dataset. Anyway, no issue with that.
But my question is this.
With prediction, we are getting a result like this:
So, how fast.ai converts those activations into a Point?
Where should I look for more info on this?
fastai scales the points to a % from -100 to +100% from the center of the image for both x and y. It (for a time) was one of the only libraries to do so (and still is). The reasoning is now any image augmentation can be applied to the points.
You’d probably want to look at TensorPoint’s encodes and decodes to see the actual conversion
Thanks a lot. That’s what I want to here.
did you manage to make this work? trying to do the same and also can’t get accurate results with images off the internet. copied your code for rendering of the dot but bot sure if that is correct?
ah wait I think I found the issue. The model will resize any input image to (240,320) so before plotting the point one needs to scale the image to that size (or alternatively, rescale the point to fit the image actual size.
ok even after fixing the rescaling mistake, results are not really great, @arunoda did you manage to get decent results in the end?