PRBonn
representation question of SemanticKITTI
Hello guys,
thanks for the effort to be able to use the KITTI data set as a semantic data set.
I have a question regarding color map, color coding. I plan to transform the LiDAR data into a BEV perspective. Common approaches use height, intensity and density of the scans and encode them in a color range of 0-255 (RGB). However, I would like to know how to maintain the typical color coding of the objects, e.g. car - blue, road - purple, wall - yellow etc. in the BEV perspective.
Thanks in advance,
Petros
Not sure what you mean; I would instinctively simply use a [0, ..., K] mapping for the K classes + unlabeled classes to keep the point cloud information.
But probably, you want to have something like a one-hot encoding. thus making a tensor $H \times W \times 3 \times K$, where the last dimension is the one-hot encoding and maybe the predictions of a network such that you get the class via an argmax over the last dimension?
i realized later, that for my approach I need GT-labels instead of pixelwise annotated data. The dataset can't provide me 3D annotations, which are required for a BEV transformation.