The TF of the camera is wrong, it seems the camera is looking done, look how the ground is not aligned with the ground plane:
The reason why the walls look thick is because the ZED depth estimation interpolate a lot in areas in image where there is no texture. Look how the wall on the left and the ceiling are deformed:
You may give a try with rtabmap subscribing to left and right images instead (stereo mode), so that depth estimation is done with cv::StereoBM instead, which doesn't interpolate textureless areas in the image (only edges). There could be also some parameters tuning that could be done on ZED camera driver to make it interpolate less the textureless areas.
There are both stereo and rgbd images in the database. How did you launch rtabmap? If you are using "rgbd_image", do you have two nodes publishing on rgbd_image? Also, the camera is still tilted toward the ground. I suggest to subscribe to point cloud of the zed in RVIZ and adjust your TF so that the cloud is aligned with the ground plane. This will help to get an occupancy grid map matching the environment.
Using the stereo mode, the cloud will be more sparse, as only edges would be estimated.