Having a hard time getting the right output as explained in the paper.

[DHDanielSuh]

Hello.
I really enjoyed reading your paper and am so excited to test your code.

Here are some outputs that used your pretrained model, both Structure3D and NYU303.

I am having a hard time getting the right output as explained in the paper.
It was happening for both pretrained models.

• Do you have any restrictions on input images by pretrained models?
• Two models is outputting different results. Which model do you recommend more?
• Do you have any restrictions on the environment settings?

Could you please share some more explanations on how to use the pretrained model you provided?
And it would be great if you can share some sample images you used for the inference.

Thank you.

[CYang0515]

Hi,
which dataset do you use separately in the three case? And the results of the first two lines seem to be correct?

[DHDanielSuh]

Hi.
I just randomly collected these images online. These images didn't belong to any dataset. In terms of the result, both first and second image wasn't outputting correct layout lines. If you see the first image, the sidewall is totally off. And for the second image, the front wall is off too.
I have also tried with images in the NYU303 dataset, but it wasn't working correctly.

• Is there any specific dataset that I have to use to do some inference?
• Is there any restrictions on input images?
• Could you please walk me through how to follow the paper's result?

Thank you.

Here are some more inference images that I tried.
Structure3D_pretrained.pt OUTPUT

NYU303_pretrained.pt OUTPUT

Structure3D_pretrained.pt OUTPUT

NYU303_pretrained.pt OUTPUT

[CYang0515]

You should do some inference on Structured3D and NYU303 test dataset. For randomly collected images online, maybe the predicted 3D parameters of planes are incorrect as the different intric metric and the data bias.

[marcomiglionico94]

I tried too and my results looks even worse than what you are showing here. I tried both the models on random images collected online and on some examples from the NYU303 dataset but none of them seem to work good. Have you done any progress?

[Mps24-7uk]

@CYang0515 I am facing the same issue, the layout inclination are going off. Can you walk us through how to replicate the same result?