search

sconlyshootery / FeatDepth

This is the offical codes for the methods described in the "Feature-metric Loss for Self-supervised Learning of Depth and Egomotion".
MIT License
247 stars 28 forks source link

feature-metric loss only use the first output of the Autoencoder #86

Closed EchoTHChen closed 2 years ago

EchoTHChen commented 2 years ago

I found in your code that the feature-metric loss only used the first output of the feature extractor(resnet) which meant it only adopted the output of the 7x7 convolution with stride-2. I wonder why not use the last output of the encoder which has a larger receptive field. in line 193 of mono_fm/net.py: src_f = self.extractor(img)[0]

sconlyshootery commented 2 years ago

The output of the 7x7 convolution with stride-2 is spatially finer than the last output of the encoder, which is too small.

EchoTHChen commented 2 years ago

Thanks!