Closed LalitAttarde98 closed 3 weeks ago
Thank you for your interest in our work! No problem at all. The lengths of the major and minor axes of the ellipse are determined by the standard deviation in the x, y direction around a map vertex (not every pixel). We calculate this for all the map vertices of a map element and plot it. Since we are trying to quantify the uncertainty around map vertices, the size of the ellipse basically represents the uncertain region.
Thanks for replying. I understand following points. Correct me if something is wrong
You're conducting image segmentation on bird's-eye view (BEV) features. The predicted lane pixels are treated as map vertices, which are then used to compute regression uncertainty. The regression is carried out using the coordinates of these vertices, rather than their pixel intensity or probability.
Actually there is no segmentation conducted in this case. The map here is vectorized, not rasterized, so the map vertices are just 2D points in the ego coordinate system. So there is nothing related to pixel intensity here.
How are the lengths of the major and minor axes of an ellipse drawn around a lane vertex determined ?
In image segmentation, do you calculate two standard deviations for predictions at each pixel ?
I apologize for the basic question. I read the paper, but I couldn't understand how you are obtaining the standard deviation at each pixel in both the x and y directions.