Got it - I think i also use some parameters which are interjecting with --scan, but I got it to work on the GUI (somewhat, the output isnt really any good as we'd likely expect from non lidar)

Quick follow up question - when lidar is disabled on the iOS app, RTABMap wont look at the depth maps (assuming there's none), is that it?

If I were to use a different model for depth estimation from the RGB frame and insert them into the depth column (matching the colour space and res expected), would I need to do anything else for the mesh to process like a LIDAR enabled scan? You mention the scan column is used in any case, this means this is calculate by ARKit/LIDAR already or processed from the depth maps directly?

Hope this make sense, as always thank you for your help!

Did some further reading and found this issue which was very helpful in understanding the derserialization process #962

Looking at the PC in the Desktop app it really looks like some missing coordinate/transform issue

Ive uploaded the DB here in case you think of something very obvious (the depth maps have been inferred using Depth Anything v2)

The script does something like
• Generates AI depth maps using Depth Anything V2 model from stored RGB images
• Converts depth estimated to RGBA depth images with 16-bit encoding (mimics the compression happening on device)
• Uses per-node camera parameters - handles varying intrinsics due to auto-focus/rotation
• Some scaling/filtering to keep a relatively short range up to 4.5 m

Here an example to reconstruct the mesh on Desktop app for scans taken without ARKit's Lidar depth image:

For the depth generated by Depth Anything V2, it seems not working that well with those images, here is an example of a corner of the room, but the two walls and ceiling appear all on the same plane:

For you previous question: "Quick follow up question - when lidar is disabled on the iOS app, RTABMap wont look at the depth maps (assuming there's none), is that it?"

• The depth image will be empty. On iOS, the "LiDAR" data is saved as depth image when it is there, the scan field contains only the raw 3D tracked points from ARKit (mainly there for debugging purpose, though can be useful for mesh rendering without lidar). So you can indeed regenerate the depth image yourself. I don't know if you can provide some pixels with known depth to help the depth estimation model, if so, you could use the "scan" points to get depth on some pixels.

This is great thanks for your help!

Ill keep digging re depth estimation and will report back if there's improvement