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
246 stars 28 forks source link

_make_crp #55

Closed wangcong607 closed 3 years ago

wangcong607 commented 3 years ago

Hello, First，thanks for your contributions. But I have a question about the function of _make_crp in your depth_deocoder.
In the code, you use the sequence of four 1*1 convolutions, so I want to know why? Thanks for your reply.

sconlyshootery commented 3 years ago

The architecture is modified from a lightweight structure used in segmentation named "Light-weight RefineNet". 1 1 convs is used to save memory, you can extend it to 3 3 convs will obtain better performance.