question about pretrained model (Resnet)

fangchangma / sparse-to-dense

ICRA 2018 "Sparse-to-Dense: Depth Prediction from Sparse Depth Samples and a Single Image" (Torch Implementation)

Other

417 stars 95 forks source link

question about pretrained model (Resnet) #15

Closed polar99 closed 5 years ago

polar99 commented 5 years ago

Thanks for releasing the source code! Your work is excellent! I have run your code (using default settings) these days on nyudepthv2. But I found that I got the best result delta<1.25 =0.64 (in paper it is 0.81). I did not use the pretrained model of ResNet. Is it the reason? @fangchangma

fangchangma commented 5 years ago

Hi. Yes, not using the pretrained weights leads to suboptimal performance.

polar99 commented 5 years ago

Thank you! I run the code with pretrained model and got a better performance (delta<1.25 =0.8093) which is very closed to the best results.