Giters

TUI-NICR / ESANet

ESANet: Efficient RGB-D Semantic Segmentation for Indoor Scene Analysis

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

third dimension issue with our images when using Cityscapes weight

Shiv1143 opened this issue · comments

commented

While using cityscapes weight,I faced errors with regarding to third dimension of our image related to encoder features in model.py .so I manipulated the third dimension by slicing it to the appropriate dimension to make it work according to our images dimension.can you solve that issue from your end

commented

Thanks for reporting any issue. Can you add a link that points to the exact line?

commented

sorry I wasn' t aware that someone has replied,.. help me in creating the pull request for this issue

commented

I still do not understand the problem exactly. Can you give us more details and point to the line in the code?

commented

In model.py file in forward function:
def forward(self, decoder_features, encoder_features):
out = self.conv3x3(decoder_features)
out = self.decoder_blocks(out)

    if self.training:
        out_side = self.side_output(out)
    else:
        out_side = None

    out = self.upsample(out)

    if self.encoder_decoder_fusion == 'add':
        out += encoder_features

    return out, out_side

Here out is not coming correctly in cityscapes so I made this change which I think will work across every weights:
def forward(self, decoder_features, encoder_features):
out = self.conv3x3(decoder_features)
out = self.decoder_blocks(out)

    if self.training:
        out_side = self.side_output(out)
    else:
        out_side = None

    **out = self.upsample(out)
    if(out.shape != encoder_features.shape):
        out = out[:,:,:,:encoder_features.shape[3]]**
    if self.encoder_decoder_fusion == 'add':
        out += encoder_features
    return out, out_side

you can verify from your side as well.

commented

and is there a way where we can raise pull request for this issue?

commented

Hmm, ... actually, this should not happen - can you share the the full python command you execute including all arguments as well as some details about your python environment (pip list, conda list)?

commented

I generally use Conda enviornment. Actually I run it quite long back so can't remember the command but the issue was that while running cityscapes weights the 3rd dimension isn't aligned with the encoder shape and have more length than expected so I modified to make it work.

commented

I guess, you just picked the wrong context module and/or input resolution - both are different for Cityscapes compared to NYUv2/SUNRGB-D.
I will close this issue as we are not able to reproduce this issue (anymore).