Padding side might be wrong about fcrn-depthprediction HOT 2 CLOSED

Hello,

that's a good observation and you are perfectly right. Actually we directly trained the NYU model with the interleaved convolutions, in which case the order of the asymmetric padding does not matter. There we set it arbitrarily, to be honest.

However, when converting the model from the slower 5x5 version to the faster one, the padded side matters to get equivalent results. Our Make3D model was only trained in the 5x5 conv setup and consecutively converted using the correct padding (as you point out too).

Thanks for your comment!

from fcrn-depthprediction.

Thanks for the clarification.

Cheers.

from fcrn-depthprediction.