anshulpaigwar / gndnet Goto Github PK

Thanks for the repo!

I wonder if you use the remove_outliers function to remove all the other points than "ground", would you just calculate the recall value?
In other words, if you just classify all points as "ground", the recall metric here will be 100%.

Can you please clarify this?

Could you please share the structure of the data folders?

I found the validation frames from the dataset you provide is not 3040, (seq_001, seq_005, seq_007, seq_009), I'am confused now.

The indexes for validation based on the preprocessed dataset do not add up to 3040.

Are there some frames skipped during evaluation or are there any differnet npys during evaluation. Please do let me know.

Thanks,
Prashant

Hi,
Thanks for releasing the code for the paper.

I wanted to know the details of the psueod image. Is it a range image form of the lidar. What would be its dimensions, and how do I interpret this psuedo-image.
I wanted to know some details regarding the Convultional Enc-Dec output. Is the output having the same dimesnison as the input lidar scan. Is the output of the decoder similar to a recosntruction of the lidar scan that was input to the models.

Eagerly awaiting for a reply.

Thanks,
Prashant

Hi, I would like to test this with a 3D lidar installed on a mobile robot.

How can I do so? Could you please provide the "entry point" for the real time data?

and the output as well?

Thank you

Hi there! I was searching for ground detection / height detection methods in autonomous driving for inspiration to develop cliff detection models. It seems like your model can detect whatever is above the ground but I was wondering whether your code can also perform ground depression (the opposite of elevation) and detect objects below the ground, like cliffs, steps and stairs (downwards). Thank you!

Could you please add np2ros_pub_2 into utils/utils.py

Hi,

thanks for providing your code to your paper. I checked point_cloud_ops.py where the point pillars are created from a point cloud. In the points_to_voxel*_kernel functions you use ndim=3. To me this indicates that you are mapping the points actually to 3d voxel cells instead of point pillars since if you have two points A=(x,y,0), B=(x,y,20) I would have assumed they are located in the same point pillar as they are just on top of each other, but for a voxel size in z < 20 they will never be associated to the same pillar.

Could you please clearify why you used ndim=3 instead of only considering x and y position? Maybe I am overlooking something.

Thanks a lot in advance.