4 ms·

I'm intrigued by the rotated schoolbus images. I wonder if the failure of the computer is due to the method of presentation. The schoolbus images are fundamenta

I'm intrigued by the rotated schoolbus images. I wonder if the failure of the computer is due to the method of presentation. The schoolbus images are fundamentally two different things: an object, and a background. Humans know this intuitively: we understand the rules of 3D reality and extrapolate them unconsciously to all the objects therein. A computer is presented with just one thing: an array of pixels. Can we really expect a neural network to understand depth, scale, motion, and rotation when all it knows is two dimensions?

mr_toad 8y ago
> Can we really expect a neural network to understand depth, scale, motion, and rotation when all it knows is two dimensions? Frozen images at that. I’d expect better results from systems trained on videos of moving objects.