Annotation cost for 1‰ label. For a dataset with dense points

[Cheater-Hater]

Thanks for your awesome work! I'd like to ask about the annotation cost of your work

For example, one of the rooms in S3DIS has over 9,000,000 points. Does that mean that you're going to annotate 9000 points for that scene?

Point clouds with that many points are not because it is large, it's just because that the points are really dense.

I'd like to know can your method still work when a dataset's point cloud is not that dense?

[PointCloudYC]

For example, one of the rooms in S3DIS has over 9,000,000 points. Does that mean that you're going to annotate 9000 points for that scene?

Yes, it is true. But you need to annotate this 9000 ( 9,000,000*0.001) points over the sub-sampled point cloud obtained through grid down-sampling.

You might refer to our unofficial TensorFlow re-implementation SQN_tensorflow.

As for the dense issue, I think it is dataset-dependent. The SQN paper had been benchmarked on 7 datasets w. different characteristics, some of them might be not that dense, e.g., Toronto3D and DALES ( acquired through aerial/mobile laser scanners) should be less dense than S3DIS (through terrestrial laser scanners).