Question about PointRcnn

[Len-Li]

Hi Xiangyu, Firstly, thanks for your great work on domain adaptation. Secondly, after reading the paper, i come up 2 question(if i have wrong understanding please point it out): 1.The idea seems easily understood: changing the car size greatly improves AP. I think this is because the detector has overfitten car size in train-set. When training PointRcnn, it use average car size of train-set as templates and then regress residual. So it's obvious that AP falls when eval on annother dataset. Do you have experiment using anchor based method like AVOD? 2.After transform other dataset to kitti format, Is it only object within 90°front view be detected? Finally, I appreciate for your work on 3d object detection. Many thanks.