keon / speed Goto Github PK

I solved this with a simple VGG feature extractor (at the early layers) followed by a linear regressor. The motivation is that 1) dataset is small and 2) different than imagenet, so you use the early layers as low level generic feature detectors and then feed those to a linear regression model.

I got around 1.19 MSE with just 30% of the dataset and down-sampled frames. You can probably get that to below 1 but might be prone to overfitting.

Hello, Your code is enormously helpful to me. Really Thank you.

BTW, i have one question about crop.
I'm also trying to crop images. but I don't know where the cropping part of your code is.
Can you explain it to me? How did you crop it?

Really Thank you :)