Usage of Normals about heat

[woodfrog]

Hi @coldmanck , thanks for your interest in this work and let me try to answer your questions one by one:

1. The intuition behind the usage of normals. First of all, I need to clarify that the data preprocessing scripts for Structured3D (your first snippet) come from the paper MonteFloor, and I had included relevant information in the README file. The normals might not help corner detection, but since points on the same flat plane (e.g., wall, floor, ceiling) shares the same normal, the use of normals could benefit edge-level or room-level estimation. There is a visualization of the normal map in my earlier work Floor-SP, see Figure 5 for the density+normal map. Note that Floor-SP and MonteFloor rely on room instance segmentation for further optimization, and normals can likely help this initial step.

2. The first snippet. For the choice of the sampling hyper-parameter, I believe it should depend on the data -- If you found that using 10 makes the density map too sparse and the structures less recognizable, you should decrease this value. The clipping into [0, 1] is indeed weird, and I cannot get the motivation behind this operation, or maybe it's a typo from the original author.

3. The second snippet. The maximum operation is just my lazy trick to make sure that the combination of the density and normal map still has three channels and is a valid image so that the pre-trained ResNet can be directly used. There is no deep intuition behind it. You can also do a concatenation and add an extra linear projection to convert the input back to three channels.

Hope it helps!