search

baegwangbin / surface_normal_uncertainty

[ICCV 2021 Oral] Estimating and Exploiting the Aleatoric Uncertainty in Surface Normal Estimation
MIT License
225 stars 22 forks source link

Training process #2

Closed Runsong123 closed 2 years ago

Runsong123 commented 2 years ago

Hi Gwangbin,

Thanks for sharing your great work! I am curious about the training process, is the network trained by end-to-end training?

Best wishes, Runsong

baegwangbin commented 2 years ago

Hi Runsong,

The network generates three intermediate outputs (H/8 x W/8, H/4 x W/4, H/2 x W/2) before producing the final full-resolution output. The training loss is applied to all four predictions (hence not end-to-end). For the coarsest prediction (H/8 x W/8), the loss is applied to all pixels. For others, the loss is computed only for the pixels selected via the proposed uncertainty-guided sampling.

Regards, Gwangbin

Runsong123 commented 2 years ago

Got it! Thanks for your quick reply! :)

Best, Runsong