I definitely agree with you, but have to add that my (amateur's) take is that this paper has actually contributed a lot. Not only have they shown marked improvements in detection accuracy, but also come up with a trainable model/framework for object detection/classification using only LIDAR point clouds (i.e. without the addition of additional sensor data such as RGB overlays). Not relying on a second (or 3rd, or nth) sensor or hand-crafted features is very impactful to the field.

I'm pretty sure that marketing is the purpose of their secrecy.

As this becomes more common that will cease to be the case unless it’s especially interesting.