Question about testing on S3DIS dataset about 3d-bonet HOT 4 OPEN

hi @LiuShihHung , the question you raised also confused me when I was doing experients. Basically, the widely used "all points" does not mean all raw points of the original S3DIS dataset, but means all the sampled points (i.e., 4096 per block). From the first paper PointNet to almost all subsequent papers to date, this misleading term is invented and used.

Here's a similar discussion:
WXinlong/ASIS#10

from 3d-bonet.

Sorry, I have another two questions want to ask, our team is also studying on 3D instance segmentation now, thus encountering some problems in practice. Hope you can give some priceless advice.

First one is, in your paper appendix A, you will train SCN semantic parallelly, is it trained on cubes or whole scenes (mapped to final result in some way)?
Second question, I know that your work is evaluated on ScanNet Benchmark, how can you map sampled cube results into unsampled whole scene? (point cloud number wouldn't match)

By the way, congrats for NIPS spotlight!

from 3d-bonet.

@LiuShihHung
(1) The SCN is trained on the whole scenes as same as its released code, so each 3D point will have a predicted semantic label afterwards. You could save these predictions for block meging. 3D-BoNet doesn't need the semantic information in training.
(2) First, test all points (not sampled) of each block and then run the block merging algorithm. Basically the output of block merging is a huge 3D volume with an instance label for each voxel.
Second, query all points of the whole scene within that 3D volume, then each 3D point could get a predicted instance label. Note: the second query step gurantees all points have instance labels and the correct order of these points for online evaluation.

from 3d-bonet.

Hello,

Regarding the first question LiuShihHung had raised, I have 2 questions in the point of semantic segmentation.

(1)
The link (WXinlong/ASIS#10) says that the performance(accuracy) would be same regardless of the number of points(4096 or all the points in the block).

I agree that the accuracy would be same under the condition that 4096 points are sampled equally(random sampling).
But i wonder what if i select the specific region of block. I would choose left-bottom corner of the block and get 4096 points from that corner. (all 4096 points are from that corner)

Then I believe the 4096 points does not have global information of the block. because they are only from the left-bottom.
Or, among the 4096 points, the points which is in the center region of left-bottom corner would have okay-accuracy because they have lots of neighbor information.

On the other hand, among the 4096 points, the points which is in the outer region of left-bottom corner would suffer from accuracy degradation because they do not have enough neighbor information.

This is my assumption. What do you think about it?

(2)
Why you and the link sample points from block not room? Many code divides the room to block and sample the points.
How about sample 4096(random sampling) from room, not block? Does it drop accuracy?

from 3d-bonet.