Poor segmentation results on multirooms in the same floor

[shayan-nikoo]

Hi Jonas,

Thanks for your excellent work. I tried Mask3D on a single room of our dataset and the results were good but when we tried it on multirooms (neighbor rooms in a floor) the results are poor for example instances are segmented partially. This is very important because most of the scans contain several rooms and separating them manually for inference is a lot of work. Ofcourse this problem is understandable because the model is trained on rooms instances of S3DIS but how can we train it e.g. on S3DIS in a way that it can handle neighbor rooms? One idea is to do the cropping 6 * 6 meters on neighbor rooms and let the model see neighbor rooms.

Do you have any suggestions?