search

shawLyu / HR-Depth

[AAAI 2021] HR-Depth : High Resolution Self-Supervised Depth Estimation
MIT License
238 stars 28 forks source link

About your network architecture #2

Closed FangGet closed 3 years ago

FangGet commented 3 years ago

Hi, Xiaoyang, First thanks for open-sourcing this project, great job. There is something inconsistent between the code and your paper about the network architecture. After digging into the code, I think the ar should be like this: image which is not corresponding to your paper.

shawLyu commented 3 years ago

Hi, Fang, Thanks for following our work. Can you tell me in detail which node has problem? Because I did not know specific problems between figure you provided and ours. Maybe we have different representations for the same architecture. Welcome to discuss! Thank you!

FangGet commented 3 years ago

From my point of view, the node [40] from my picture corresponding to X_5^e from the paper, so how to generate the X_5^d from X_5^e from your paper? for my picture, I see [40] is fused into 31 with directly upsampling.

FangGet commented 3 years ago

or your X_5^e is exactly the same as X_5^d, without any operation? just for expression convenience?

shawLyu commented 3 years ago

I'm sorry for ambiguity. In our paper, we used a red arrow to connect $X_5^e$ with $X_5^e$ which means convolution with $3 \times 3$ kernel. In the code, self.convs["X_40_Conv_0"] represented for this operation. $X^5_e$ represented for features extracted from encoder corresponding to input_features[5] in the code and $X^5_d$ is the intermediate variable in nestConv function. You can check the code again and I hope this answer can solve your problem.

FangGet commented 3 years ago

ok, I will check for that, thanks for explanation.