search

TRI-ML / packnet-sfm

TRI-ML Monocular Depth Estimation Repository
https://tri-ml.github.io/packnet-sfm/
MIT License
1.23k stars 242 forks source link

Question about the reprojected distance loss function #183

Open zxiaomzxm opened 2 years ago

zxiaomzxm commented 2 years ago

Hi! I have reimplement the reprojected distance loss in my own framework, but I found it hard to converge, because my initial pose is close to identity, thus in this scenario, the reprojected loss term is close to 0, right？Thus this scenario is a local minimal for this loss, it backpropagate to the pose directly, makes the photomatic loss diverged, and I can not get a good depth map in this setting. Am I missing something?