facebookresearch/votenet

Input Features

Alphastrick opened this issue · 0 comments

Hi, I have a question about the input data.

I've developed a custom dataset, but now I wonder why the model performs so poorly.
One reason could be, that I misunderstand the input dataframe, so here are some questions:

angle_class_label and angle_residual_label: Since scannet has only axis aligned bounding boxes (does it?), how to initialize these attributes?

point_votes: I looked up how you calculate your voting points and found out, that you subtract the scanned instants points from the instance center. This rotates the point votes. Is there a reason for? (I haven't found one in your paper. Did I missed it?).