Question on SnowyKITTI

[ldkong1205]

Hi @alvariseppanen, thank you for open-sourcing your work!

I have two questions regarding this work:

• When will the SnowyKITTI data be available?
• Is SnowyKITTI sharing the same configurations (e.g., beam numbers, FOV, etc.) with SemanticKITTI? Since you mentioned that you built this dataset upon the KITTI odometry benchmark.

Thanks again and looking forward to your reply!

[alvariseppanen]

Hi,

We are currently looking through the options for publishing the dataset.
In the meantime, you can use the toy dataset or generate your own using the simulation model (https://github.com/SysCV/LiDAR_snow_sim).
Note that KITTI does not have channel numbers, so we did the projection and used row indices as channel numbers.

SnowyKITTI is sharing the same configuration with SemanticKITTI.

Best regards,
Alvari

[ldkong1205]

Hi,

We are currently looking through the options for publishing the dataset. In the meantime, you can use the toy dataset or generate your own using the simulation model (https://github.com/SysCV/LiDAR_snow_sim). Note that KITTI does not have channel numbers, so we did the projection and used row indices as channel numbers.

SnowyKITTI is sharing the same configuration with SemanticKITTI.

Best regards, Alvari

Thank you for your reply. Looking forward to the dataset!

[alvariseppanen]

Hello,

The dataset is online now
https://www.dropbox.com/s/o3r654cdzfl405d/snowyKITTI.zip?dl=0

Please cite our work, citation is coming soon (accepted by IEEE Robotics and Automation Letters).

Best regards,
Alvari

[ldkong1205]

Thank you for your information! Could you share with me what horizontal resolution you used for the projection?

[alvariseppanen]

I used 2048 horizontal resolution

[ldkong1205]

I used 2048 horizontal resolution

Thank you for your reply!

Regarding the SnowyKITTI dataset, I have the following two questions:

• I noticed from the lidar_snow_sim paper that they actually provide two situations: snow and wet ground. For your generated SnowyKITTI dataset, are you using only the snow option, or both of the options are used?
• The second one is related to the file size. I saw from DropBox that your dataset is 43GB. Have you included all raw point clouds for train, val, and test sets, as well as their semantic labels?

Thanks and look forward to your reply!

[alvariseppanen]

Hello,

Only snow is used.
Raw point clouds and snow labels are included

[eugeniu1994]

Note that KITTI does not have channel numbers, so we did the projection and used row indices as channel numbers.

@ldkong1205
Hi, thank you for sharing your work.

Is there a way to get the ring/channel information for the snowKITTI dataset?

Since the point cloud in the original KITTI dataset is ordered the ring/channel can be inferred via:
int previous_quadrant = 0, ring = 0, total_points = pointcloud_in.size() - 1;
for (int i = total_points; i >= 0; i--)
{
int quadrant = get_quadrant(pointcloud_in[i]);
if (quadrant == 4 && previous_quadrant == 1 && ring < 64 - 1)
{
ring += 1;
}
previous_quadrant = quadrant;
pointcloud_in[i].ring = ring;
}

However, the same does not work for snowKITTI, since there are corrupted/snow points added, and also the number of points in each original KITTI scan is different, e.g., for snowKITTI the number of points is always the same

Regards!