Yes, the pose is with respect to the initial position. With the small bonus that on startup it uses the imu to get where down actually is so orientation "should" always be relative to gravity.
It does look like it's integrating its position.
When everything is fine, drift seem to be quite good (I haven't tried it over long distances though) but the odometry can get lost (I have tested it in my office which has some plain white walls) and then things get a bit weird: the pose can drift over a few meters before it stops (the slamdunk detects it's lost) the translation is blocked until odometry can be restarted (usually by stopping and looking at someplace with more features). But orientation doesn't stop in this phase and position doesn't reset to 0 after that. So drift can accumulate in bursts. But keep in mind that I did test it in a far from ideal environment.
I haven't tested it on a mobile robot, only holding it in hand but I've done a bit of SLAM with it (using the RTAB-Map package) and, after I found out how to handle the thing getting lost and as long as it's not doing it too often it works well.