search

TRI-ML / packnet-sfm

TRI-ML Monocular Depth Estimation Repository
https://tri-ml.github.io/packnet-sfm/
MIT License
1.21k stars 241 forks source link

Full Surround Monodepth from Multiple Cameras, I have some questions about this paper，Thank you for your answer！ #204

Open abing222 opened 2 years ago

abing222 commented 2 years ago

1、In multi camera training, do you only use spatio-temporal loss or temporal + spatio-temporal loss? 2、In multi camera training, do you use COLMAP depthmap supervise？

abing222 commented 2 years ago

"By introducing spatial contexts (Equation 3), we not only further improve performance, but also learn scale-aware models by leveraging the camera extrinsics", What does scale-aware mean?

VitorGuizilini-TRI commented 2 years ago

1) We use spatial + temporal + spatio-temporal 2) We do not use COLMAP supervision

Scale-aware means that the pointclouds are metrically accurate (i.e. estimates are in meters).

I hope that helps!

abing222 commented 2 years ago

thank you very much，In my experiment，without additional supervision, due to the introduction of camera external parameters, the scales of space and temporal are inconsistent (absolute scale and relative scale), resulting in poor model training results