search

zzangjinsun / NLSPN_ECCV20

Park et al., Non-Local Spatial Propagation Network for Depth Completion, ECCV, 2020
MIT License
321 stars 57 forks source link

Just multiply confidence and initial depth together while propagating! #52

Closed XJTUXYC closed 1 year ago

yudmoe commented 1 year ago

Hi bro, I have the same question about why the confidence should be propagated, the function ModulatedDeformConvFunction looks quite complex, . Did you get the result of just multiply confidence and initial depth?

zzangjinsun commented 1 year ago

Hi, you SHOULD NOT multiply confidence directly to depth values because confidence-multiplied affinity values are normalized and the normalized values are multiplied to depth values.