As I mentioned, can you get a depth buffer from the camera? If so it’s just math inside a post-process shader:
Lerp(PostProcess0, CameraFeed, clamp((SceneDepth - CameraDepth) / TransitionDistance));
Given that ARKit plots the surfaces in general I’m sort of surprised that it isn’t building some kind of occlusion mesh to do this automatically. If so it might live in (or can be pushed into) the custom depth buffer, which would be even better than the camera depth feed if such a thing exists.