Thank you. You did not implement the 3D parameters Construction?

[youjianghuang]

Why?

[youjianghuang]

and how to?

[sunset1995]

Hi @youjianghuang
I assume you mean "3D Layout Parameter Regressor" branch in their network architectures.
According to LayoutNet paper, with or without that branch seem trivial in their framework:

• "We observed that the 3D regressor is not accurate"
• "Training with 3D regressor has a small impact"

To implement it, you probably need:

1. Implement a regression network which take B x 4 x 512 x 1024 tensor as input and output B x 6 3D parameters
2. Concat y_edg_ and y_cor_ then feed it to regression network
3. Prepare 3D parameters dataset from official ground truth data