does it has the detail in the paper ?

[vicentowang]

In PackResNetEncoder x = (input_image - 0.45) / 0.225 image norm? why i don't see it in the paper?　and more details　I　cannot find it from the paper.

[FangGet]

Nope, they did not provide detail for that. This project is essentially based on monodepth2, so many details and trick are adapted from it. btw, for PackResNetEncoder x = (input_image - 0.45) / 0.225, it's an input normalization to get numeric stable.

[vicentowang]

there still exist many conflicts with the paper , as I only read the encoder parts ,two issues: 1.BasicBlock
refer to the ResidualBlock ,i think, this should be conv -> ELU -> conv ->ELU -> conv-> ELU -> GroupNorm -> dropout 2. for i in range(1, blocks) layers.append(block(in_channels, out_channels)) ,cascade-ResidualBlock not mentioned in the paper. In my opinion though.

[FangGet]

yeah basic residual block is not the same which after testing not influence the result.

for cascade residual block, paper mentioned, which you can see at Table 1 for network detail

[vicentowang]

yeah basic residual block is not the same which after testing not influence the result.

for cascade residual block, paper mentioned, which you can see at Table 1 for network detail

cascade residual block, paper mentioned　table1 you mean this ? Each ResidualBlock is a sequence of 3 2D convolutional layers with K = 3/3/1 and ELU non-linearities, followed by GroupNorm with G = 16 and Dropout [41] of 0.5 in the final layer. I can't see from it. more detail ?

[FangGet]

x2 means duplicate ResidualBlock 2 times.

[vicentowang]

thanks. got it

------------------ 原始邮件 ------------------ 发件人: "Fangget"<notifications@github.com>; 发送时间: 2020年1月16日(星期四) 晚上6:38 收件人: "FangGet/PackNet-SFM-PyTorch"<PackNet-SFM-PyTorch@noreply.github.com>; 抄送: "渚薰岚"<849529792@qq.com>; "Author"<author@noreply.github.com>; 主题: Re: [FangGet/PackNet-SFM-PyTorch] does it has the detail in the paper ? (#2)

x2 means duplicate ResidualBlock 2 times.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.