I think the code and the contents of the paper are different.
jaeseok4104 opened this issue · 6 comments
jaeseok4104 commented
samsdolphin commented
Hi @jaeseok4104, yes you are right. In short, we found this implementation has a faster convergence rate and a better precision. We thus open source this version of code instead of the very original one (paper version). However, you could modify the code structure by yourself. Hope this helps.
jaeseok4104 commented
thanks :) Have you plan to open about paper version code?
samsdolphin commented
Haha, this shouldn't be hard for you ;)
HaisenbergPeng commented
samsdolphin commented
Hi @HaisenbergPeng, in each local ba, we want the hessians to constrain the relative poses between the 1st layer poses. The optimized poses in each local ba are used to construct the next layer's keyframe.
HaisenbergPeng commented
okay, got it! Maybe I could try using the optimized poses of the bottom layer as factors.