Hi Francesco,
Do you reproduce the results of the point pillars paper on kitti dataset (I have not seen some metrics, like mAP) ?
In a first time, I have implemented on the side a non maximum suppression after the network to compare the predicted box and GT. But I have a doubt on your angle implementation. Like in second paper, I have tried Yawlabel - YanAnchor in ground truth generation and then something like soomthL1(sin(Ypred - Ygt)) but the network doesn't seem to be able to predict angles at all.
Do you test it on your side ?
best regards
Pierre
Hi Francesco,
Do you reproduce the results of the point pillars paper on kitti dataset (I have not seen some metrics, like mAP) ?
In a first time, I have implemented on the side a non maximum suppression after the network to compare the predicted box and GT. But I have a doubt on your angle implementation. Like in second paper, I have tried Yawlabel - YanAnchor in ground truth generation and then something like soomthL1(sin(Ypred - Ygt)) but the network doesn't seem to be able to predict angles at all.
Do you test it on your side ?
best regards
Pierre