The generated kitti box position information has deviations
wanghui-min opened this issue · comments
After importing the original kitti data, you will find that the boxes are all deviated downwards.
Using labelcloud to mark up the kitti format of the label into open3d and other software to open also exists up and down deviation.
Hi @wanghui-min,
are you sure the bounding box coordinates are in the coordinate system of the point cloud and not in the camera coordinate system?
Maybe have a look at this closed issue: #35
You can also try the kitti_untransformed label format.
Hi @wanghui-min,
are you sure the bounding box coordinates are in the coordinate system of the point cloud and not in the camera coordinate system?
Maybe have a look at this closed issue: #35
You can also try the
kitti_untransformedlabel format.
Previously, because of the labeling after using deep learning model for training and detection found that the prediction box will be biased, after the labelcloud original generated label using other tools to display found that there are errors, so the label results into the camera coordinate system projection to the image, found that there are errors in the label, so I suspect that the problem is generated by the labeling time (because I did the point cloud and RGB image fusion, camera and radar conversion matrix is no problem). So the original kitti data was replaced by the radar coordinate system using labelcloud display, and the deviation problem also appeared.
Can you provide exemplary point cloud files and their labels?
Can you provide exemplary point cloud files and their labels?
data.zip
There are three in there, all from the kitti data set, the 22-24 data in the original training set or test set.
Hi @wanghui-min,
did you solve the problem or why did you close the ticket?
Hi @wanghui-min, did you solve the problem or why did you close the ticket?
I found what seems to be a center point problem. The center point generated by labelcloud is the center of the cube, and the network model seems to have the input as the center of the bottom surface of the cube. When I converted the center point generated by the labelcloud box to the center point of the bottom side of the cube box, I found that the position was correct, but the kitti data loaded in still showed a downward shift. I think the problem is caused by the fact that the kitti data I loaded is also centered on the bottom surface, while the labelcloud is centered on the cube (probably because I had a problem setting the center point when converting the kitti from the camera coordinate system to the radar coordinate system).
@ch-sa Hello,I want to know whether the three-digit coordinates of an object (x, y, z) in a kitti format annotation exported using a labelcloud annotation indicate whether the object's centroid coordinates under the radar coordinate system or the center coordinates of the base surface.