Results :-

The velocity estimated from AB3DMOT with Monoflex is compared to velocity estimated using AB3DMOT with PointRCNN. The MAE decreases by 0.03 and can be seen in the following table 3.

Discussion :-

This is a simple non-learning based algorithm which is heavily dependent of the performance of detector. The 22 keypoints approach is combined which gives a velocity MAE of 0.1 tested only on KITTI validation dataset. This is pretty accurate given that it is a non-learning based method and has very high inference speed.