benchun123 / point-plane-object-slam Goto Github PK

hello benchun,
I think your work has a wonderful performance, but when reading your article, I have a question about the 2d bbox regression step, i.e. how can we get an object's 2D bbox on the ground plane from a rgb image?
hope for your reply:) thanks

Hi,Benchun, thank you for your excellent work. I would like to ask when the code is open source?

Hi @benchun123,

Once again, thank you very much for open-sourcing your excellent work.

Can you clarify what you meant by "In this code, we read offline 3D cuboid, the objects can be detected online."

I understood the latter part of the statement i.e. pass bounding box and class label during runtime but why would you need to read the 3d cuboids offline? Aren't you generating them during run time?

With best,
@Mechazo11

Hello @benchun123

In your paper, how is $X_{3D}$ computed? After Equation 1 it appears $X_{3D}$ points are no longer used.

Are you using vanishing point methods similar to how CubeSLAM first find 8 corners in 2d plane and then use PnP to project them to 3D space to find the 3d position of the 8 corners of the cuboid?

Also is $P_{roj}$ equal to $K$ the calibration matrix?

Hello @benchun123

Thank you very much for making your code open source and the paper publicly available. I am trying to compare your method against a cuboid detection strategy that I am currently working on.

In section 4.1.2, the paper states the following "To compare our proposed detection method with SUN Primitive [35] and CubeSLAM [20], we select 400 images with visible ground planes and fully visible ground objects"

May I have a copy of those 400 images so that I can test the 3D IoU approach with Structure SLAM for appropriate comparison?

With best,
@Mechazo11