Compared with ROS's compressed_depth_image_transport, we add an additional signature, but there is no compressionConfig part. Therefore, additional processing is required to convert from compressedDepth message to RTAB-Map.

https://github.com/ros-perception/image_transport_plugins/blob/ccb2e9bf447c16c1857afc0d935413dcc2969cea/compressed_depth_image_transport/src/codec.cpp#L383-L392

If someone uses rtabmap_util/rgbd_relay for data compression and then rtabmap_util/rgbd_split for splitting, they will not get a valid compressedDepth message.

That PR looks great! I'll check tomorrow how it could work with ROS. As you said, we may have to strip the header "DEPTHRVL" from the byte array when converting back to sensor_msgs/Image. I guess for the input compressed depth images, image_transport would already convert it.

On rtabmap_ros, we don't actually convert back from rtabmap's internal compressed depth format to the outside. For rgbd_split, rgbd_relay and rgbd_sync nodes, they already explicitly use PNG by default when compress option is enabled (I created this follow-up issue on rtabmap_ros). So we should not break anything merging this PR.

I did a small comparison of the compression time using ZED 2K depth images:

• RVL ~12 ms
• PNG 16bits: ~24 ms
• PNG 32 bits: ~50 ms

My main changes are:

• In DBViewer, there is an action to edit the depth image. I added a function to read the compression format so we can re-save the edited depth to database with same format.
• Because default of Mem/SaveDepth16Format=false and Mem/DepthCompressionFormat=".rvl", I removed the warning about setting Mem/SaveDepth16Format to true if Mem/DepthCompressionFormat=".rvl" (as it would appear every time). If the input depth images are already 16bits, no warning will be shown and RVL is used. If the input images are 32bits, there will be a warning if RVL is used that depth images are incompatible and they will be saved in PNG instead. This will make sure that stereo depth images, that have often values over 65 meters, are not clipped by default.