search

lmb-freiburg / flownet2

FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks
https://lmb.informatik.uni-freiburg.de/Publications/2017/IMKDB17/
Other
1k stars 318 forks source link

why the flow_loss2 is too large? #39

Closed Lvhhhh closed 7 years ago

Lvhhhh commented 7 years ago

when i train the C network. and i set the l1_loss_param: l1_loss_param { l2_per_location: false normalize_by_num_entries: true } the flow_loss2 is lager than other losses even though the settings of these layers are similar.. i check the ptototxt many times and cant find any problems so come here for help.

Lvhhhh commented 7 years ago

btw, i fond the CS network. why using the predict_flow2 as the input of warp.? Is it better to use the predict_flow1 as the input of the warp layers? does this have something with the problem 1?

Lvhhhh commented 7 years ago

i will appreciate if you can answer my question!

Lvhhhh commented 7 years ago

as the time going on . the value of flow_loss2 is similar to the other loss.....and i am confused!!