Open claragarciamoll opened 2 years ago
To remove duplicated walls, you can adjust the length parameter of quad NMS in line 22, ap_helper_pq.py. And we're sorry that our method performs not so well in some cases, like short walls or low walls as you showed, since most walls in scannet trainset are wide and high and aligned with canonical directions. This may be solved with a more balanced layout dataset that has more corner cases.
Hello, We have followed the steps on the Readme file and the quantitative results achieved (for train and validation datasets) were the same as the ones in the paper (Bounding Box detection = 67.2; Layout detection = 57.9). However, the qualitative results on the quad stage were not as good as expected (following the paper). In some cases, the algorithm detects walls where they should not, while, in some point clouds where there are points enough, the layout of some walls is not detected. I will attach some examples of these aspects next.
In scene 00642_00: There are many walls with similar normal and only one is expected.
In scene 00645_00: Some walls do not achieve a layout, and they seem to contain point enough. While the diagonal layout is detected as a False Positive.
In scene 00667_00: The same here, where the wall contains a lot of points, but the model is not able to create the layout.
I would like to know if there is any parameter that improves these results. Thanks in advance.