lfz

it's the logit before softmax the softmax function is put in the loss.py

register kaggle and attend the competition, you will find the data download path

I don't think it's normal, what is your batch size?

The labelmapping function is very similar with the Faster-RCNN, please read that paper for the explanation of anchor the five columns stand for [probability, dx, dy, dz, dr], you can...

@totesarana that's correct specifically, the x,y,z posision are all normalized to [-1,1]