Pillars/Obstructions

[harshilcontineu]

Hey First of all, great implementation! so thank you for that.

I was using this to create floorplans of spaces with thick rectangular pillars in the middle of it. Does this implementation account for such objects? I'm assuming not but I wanted your thoughts on it.

Thanks Harshil

[EnriqueSolarte]

Hi, Nice to hear from you.

Yes, you are right. Directly using the current solution may not deal with pillars, furniture, or any large occlusion in the middle of the rooms. The main limitation is due to the layout model because the data domain used during training differs from the scene application. You can check this 360-mlc.

You can try the following: NO TOO EASY OPTIONS:

• You can verify that the issue mentioned here is due to a domain gap. If so, you may want to search for other pre-trained weights or even other backbone models e.g., (LGTNet)[https://github.com/zhigangjiang/LGT-Net].
• Another option is self-training the model into your data w/o labels, you can check 360-mlc.

EASY OPTION:

• You could project only the closest boundaries to the camera. Generally, the obstacles in the middle of the room are far from the camera view. Moreover, the walls to the camera view are usually more stable.