questions

[Adam1904]

May i ask questions again?

1- where does the z axis value for the generated virtual points come from? It is calculated by the interpolated depth ? right?

2- virtual points are generated from the instance masks and not from Frustum? right?

The paper says: "We start by randomly sampling 2D points s 2 m from each instance mask m."

but it also says:
"we generate virtual points from each frustum Fj ."
and:
"We augment sparse Lidar point cloud with dense semantic virtual points generated from 2D detections".

I think the first sentence is the exact sentence.

3- Why is dist_thresh=3000 chosen? Based on what?

4- Will virtual points be generated even if there is only one lidar point in instance mask?

5- How to take class scores of the virtual points after entering the real and virtual points into the 3d detector? Is there a certain percentage that is taken, for example, 50% of the 2d detector score and 50% of the 3d detector score?

6- For example, if there is a mask for a distant car, one lidar point is projected within it and 50 virtual lidar points are generated with a 50% score, that's a car. How would the final result be calculated in the 3d detector?

7- how are the virtual points unprojected into 3D? where does the inverse matrix transform come from? from the neighbor lidar points from which the depth is taken?

8- What is painted points here?
https://github.com/tianweiy/CenterPoint/blob/8fcdc944bbb455bd25943b331aaf961aa0ab32cd/det3d/models/readers/dynamic_voxel_encoder.py#L28