Which methods does QuickLook employ for depth estimation or SLAM capabilities without a LiDAR scanner?

Question

I recently discovered that there is a "web-AR" feature for mobile iOS devices called QuickLook, where a pre-installed native solution is started for AR viewing of 3D models with SLAM capabilities such as surface detection, feature-tracking (anchoring of the digital object) and occlusion. Now, this worked impressively well, but the device I tested this on was an iPad Pro with a LiDAR scanner. I learned that QuickLook, which is implemented using the RealityKit/ARKit, automatically uses the LiDAR scanner if present which I assume massively enhances depth perception and accuracy.

Can anyone tell me which techniques the underlying frameworks uses to function without additional hardware such as LiDAR? I couldn't find specifics in the documentation about QuickLook. For example, does it use monocamera algorithms to estimate depth? Does it run ML models? I am aware that the utilized frameworks likely have capabilities for all of these, but I want to specifically know how QuickLook handles these hardware/device fallbacks, in which order and whether this is public information.

Answer 1

It's explained in detail how iOS does it, in Capturing Photos with Depth article.

For compatible devices with back-facing dual camera, AVFoundation framework can provide the depth information of every frame in the camera stream.

More useful information can be found in AVDepthData description.

Here is also an article on how depth map can be provided manually using CoreVideo APIs.