search

SeokjuLee / Insta-DM

Learning Monocular Depth in Dynamic Scenes via Instance-Aware Projection Consistency (AAAI 2021)
Other
220 stars 27 forks source link

ImageNet pretrain #9

Closed fengziyue closed 2 years ago

fengziyue commented 2 years ago

Hi, thank you for sharing this amazing work, I just have a little bit confusion. Does the AbsRel error of 0.112 on kitti eigen split with or without imagenet pretraining? If not, do you have results with pretraining?

Thank you

SeokjuLee commented 2 years ago

Hi, thank you for your interest! The numbers are the results of our imagenet pretrained model as described in the table in this section. I guess you might be confused by the numbers of MonoDepth2's results in Table 4 in our paper. Sorry for the confusion.

fengziyue commented 2 years ago

Thank you for clarifying this!

I have one more question, in your code, you are always using the models.DispResNet() as depth net. Is this corresponding to the fourth row (DispResNet, AbsRel 0.124) of Table 4 in your paper instead of the eleventh row (ResNet18, AbsRel 0.112)? If yes, how can I switch to the ResNet18 configuration?

Thank you!

SeokjuLee commented 2 years ago

The default configuration of this code is based on the ResNet18 architecture (AbsRel 0.112). In this line, you can see the input configuration of the ResNet types, which is declared in this line. The naming of "DispResNet" is originated from SC-SfMLearner and CC.

fengziyue commented 2 years ago

Ok, got it. Appreciate your prompt response.