ScanNet dataset about 3d-bonet HOT 6 CLOSED

@BenChenCh It's pretty easy.
(1) Use the ScanNet official code (https://github.com/ScanNet/ScanNet/tree/master/BenchmarkScripts/3d_helpers) to extract raw point clouds and the sem+ins ground truth.
(2) To divide the large point clouds into 1mx1m blocks. Refer to PointNet code (https://github.com/charlesq34/pointnet/blob/master/sem_seg/indoor3d_util.py)
(3) If you need to use SCN to train the point semantic segmentation, refer to code (https://github.com/facebookresearch/SparseConvNet/tree/master/examples/ScanNet).

Note, just remove the "psemce_loss" from the "end_2_end_loss" , if you don't use the pointnet++ based semantic segmentation branch.

from 3d-bonet.

Hi @Yang7879 Does the training process input all the points from 40 classes or just fewer points from 18 classes or 20 classes (including wall and ground). Some papers input points that belong to 18 classes, would this cause a problem during the testing process, which has to input all points?

from 3d-bonet.

Hi @519830100 In my case, the sparseCov was trained on 20 cls (same as their settings) for semantic seg; for instance segmentation, I tried either 20 cls or 40 cls for training, the results are quite similar. This is sensible cuz our 3D-BoNet doesn't require the semantic label for training.

from 3d-bonet.

@BenChenCh，hi，I am interested in trainning with scannet database, too. Have you successfully generate raw point clouds and the sem+ins ground truth? can you please tell exactly steps to extract raw point clouds and the sem+ins ground truth. thanks

from 3d-bonet.

@Yang7879 , hi, I can't understand the meaning of the sem+ins ground truth. Does it mean export_semantic_label_grid_for_evaluation? What is the meaning of "ins" ?

from 3d-bonet.

@Yang7879 @BenChenCh @bonbonjour
Hi, I have some trouble with extracting the semantic and instance labels on ScanNet_v2.
Cloud you please tell exactly steps to extract raw point clouds and the sem+ins ground truth?
Hope to hear from you soon！

from 3d-bonet.