Distortion on pointcloud about waymo-open-dataset HOT 13 CLOSED

Hi there, sorry for the delayed response.

1. For the distortion, are you talking about the near range lidar?

2. The 3d labels are in meters.

from waymo-open-dataset.

Thanks for your feedback. I cannot access you link. Could you help to fix that?

from waymo-open-dataset.

The following is the .png file @gogojjh reported. You need to copy the link instead of clicking on it.

By the way, I would like to know whether meter is the unit of 3D label(x, y, z and length, width, height). Thank you.

from waymo-open-dataset.

Thanks for your reply.

from waymo-open-dataset.

Hi peisun1115,
I think that the distortion may cause biased results on the dataset. For example, when the distortion is large, the ground-truth (bounding box) may not cover all the points of a vehicle. It would be better if the distortion can be fixed as the KITTI did.
Thanks.

from waymo-open-dataset.

Hi gogojjh,
Can you explain what do you exactly mean by 'distortion' here?
Thank you
Pei

from waymo-open-dataset.

Hi Pei,
Sorry for my late response. LiDAR distortion is usually caused by the fast motion of the vehicle. For example, please see the below pictures which I selected two cases from the dataset:

Case 1: without motion

Case 2: with fast motion

In the case 2, you can see that the point clouds of the front and rear LiDAR does not form a regular, circular structure. In other words, such distortion introduces additional noises to the point clouds. making object detection harder. This problem can be solved using pose interpolation.

from waymo-open-dataset.

Hi gogojjh@:
Thanks for your explanation on the distortion. I think this is expected as the lidar spins are captured from left to right (clockwise). The lidar spin starts from the middle of the SDC's back. I think it is better to present the user with the raw data. The user is free to run anything algorithm to compensate this effect.

Thanks
Pei

from waymo-open-dataset.

Hi gogojjh@:
A little more information:

1. The points are not collected at the same instant and the effect is strongly analogous to the rolling shutter effect in image sensors.

2. You will see similar artifacts on moving objects captured by a stationary car. The issue is relative motion between sensor and sensed object.

from waymo-open-dataset.

Hi Pei,
Thanks for your reply. This information is very helpful.

Thanks,
JIao.

from waymo-open-dataset.

@gogojjh
Can you elaborate on how KITTI fixes the distortion caused by the rolling shutter effect in fast movement lidar?

from waymo-open-dataset.

In my view, WOD has effectively addressed the distortion issue. Otherwise, you would notice that the fourth picture displays symmetrical points deployed on the flat ground surface, similar to what the second picture illustrates. I.e., the WOD point cloud more accurately reflects the actual location compared to a perfectly enclosed circle placed on a flat ground surface while the car is in motion. Is my understanding correct? @peisun1115 @gogojjh

Hi Pei, Sorry for my late response. LiDAR distortion is usually caused by the fast motion of the vehicle. For example, please see the below pictures which I selected two cases from the dataset:

Case 1: without motion

Case 2: with fast motion

In the case 2, you can see that the point clouds of the front and rear LiDAR does not form a regular, circular structure. In other words, such distortion introduces additional noises to the point clouds. making object detection harder. This problem can be solved using pose interpolation.

from waymo-open-dataset.

Hi! I want to know that is there any good method for motion object distortion? I only found a paper-'Lidar with Velocity: Correcting Moving Objects Point Cloud Distortion from Oscillating Scanning Lidars by Fusion with Camera',but it's for the horizon livox lidar and my lidar type is velodyne.

from waymo-open-dataset.